Download PDF Fall Bridge on the Materials Genome Initiative September 29, 2025 Volume 55 Issue 3 The Fall 2025 issue explores the Materials Genome Initiative’s progress and future outlook, showcasing advances in autonomous experimentation, sustainable polymers, next-generation batteries, and the broader role of AI in engineering. Accelerated Materials Discovery Through the Power of Artificial Intelligence for Energy Storage Monday, September 29, 2025 Author: Arumugam Manthiram and Tianxing Lai From minerals to machine learning: shaping the next generation of batteries. Energy storage with batteries has become an integral part of our daily life, ranging from portable electronics, such as cellphones and laptops, to electric vehicles (EVs). They have the potential to transform grid storage of electricity as well. Batteries also play a critical role in national defense, ranging from soldier power to communication devices to vehicles. The global lithium-ion battery market was projected to be valued at around $60 billion USD in 2024 and to reach ~$182 billion in 2030 (The Research Insights 2025). For a battery technology to be widely adopted, several critical parameters need to be considered: cost ($ kW–1 h–1), energy density (W h kg–1 or W h L–1), power density (W kg–1 or W L–1), charge–discharge cycle life, safety, and environmental (toxicity) impact. These parameters are linked to severe materials challenges (Figure 1). The dominant factors that need to be considered depend on the application. For instance, for portable electronics, user time between charges (energy density) is dominant because these batteries are small and cost is not an impediment. For EVs, driving range (energy density), cost, safety, cycle life, and fast charge (power density) are all critical factors, generally decreasing in priority as battery size increases. And for grid-scale storage, the dominant considerations are cost, reliability (cycle life), and safety, given the large size of the batteries. The Cost and Supply-Chain Challenge The battery market is dominated by lithium-ion batteries due to their high energy density and long shelf-life. However, they employ expensive metals like cobalt, nickel, and lithium that are naturally scarce. As the battery industry rapidly expands for transportation and grid storage applications, cost—while ensuring adequate safety—will remain the dominant concern. Raw materials availability, their processing, and cell manufacturing all contribute to cost. Other performance factors such as cycle life (durability) and safety also impact the cost. A battery with a long cycle life requires fewer replacements, lowering overall costs; similarly, strong safety performance simplifies operation and management, further reducing costs. Battery raw materials abundance and origin also impact the supply chain, national security, and national economy for the United States. For example, cobalt—which still dominates portable electronics batteries—is a critical metal that is mined largely in the Democratic Republic of Congo in Central Africa where child labor presents ethical issues. Lithium mining in South America poses environmental hazards and social concerns. Factors such as these can pose supply-chain challenges to the United States. Therefore, it is imperative for the scientific community and industry to explore and develop affordable, supply-chain-friendly battery chemistries and materials with adequate safety—particularly for EVs and grid storage—in order to relieve the United States and society in general from its dependence on critical materials like cobalt, nickel, and lithium. Artificial intelligence (AI) can play a role in this regard if appropriately integrated with physical intuition and scientific curiosity. Accordingly, this article first gives an overview of how we arrived at where we are now with battery technologies, reviewing the scientific and engineering innovations that have occurred over the past half century. Then, it focuses on how the power of AI could help accelerate the discovery of new materials and battery chemistries, highlighting the recent developments and accomplishments with AI. Energy storage with batteries has become an integral part of our daily life, ranging from portable electronics, such as cellphones and laptops, to electric vehicles. How Did We Get to Where We Are Now? Lithium-ion batteries operate on intercalation chemistry, in which lithium ions are reversibly inserted into and extracted from the anode (negative electrode) and cathode (positive electrode). The anode and cathode are separated by an electronically insulating but ionically conducting electrolyte, which transports lithium ions (the working ion) between the electrodes to maintain charge neutrality, while electrons flow through the external circuit to perform useful work. Both the anode and cathode should ideally be good electronic and lithium-ion conductors to transport electrons and working ions and support acceptable charge–discharge rates and power density. The anode should have high negative electrochemical reduction potential while the cathode should have high positive electrochemical reduction potential to maximize the operating cell voltage. In addition, both the anode and cathode should facilitate a large degree of reversible lithium intercalation/deintercalation to maximize the amount of charge stored (cell capacity). The product of cell voltage and capacity determines energy density. The reversibility of the two electrodes over many charge–discharge cycles along with their interfacial stability in contact with the electrolyte determines the cell cycle life and durability. Issues like metallic dendrite formation, internal short circuits, cell swelling, and gas evolution as the cell cycles, along with electrolyte flammability, can degrade cell safety. Toxicity associated with the materials, processing, and manufacturing determines the environmental (including human health) impact. These factors and the costs associated with them highlight the significant challenges inherent in materials design, development, and implementation. Although intercalation chemistry of ions or guest molecules into solid hosts was known for close to two centuries, Stanley Whittingham was the first to demonstrate a rechargeable lithium battery in 1976, employing lithium metal as an anode and titanium disulfide (TiS2) with a layered crystal structure as a cathode (Whittingham 1976). Direct metal-metal (Ti-Ti) interaction and the two-dimensional layered structure supported good lithium-ion conductivity. Following the demonstration with TiS2, several layered sulfides and selenides were investigated as cathodes, and rechargeable lithium batteries with metallic lithium anode and a sulfide cathode began being marketed in the 1980s. Unfortunately, dendrite growth with lithium metal and internal short-circuit led to fire hazards, resulting in an abandoning of the technology. John Goodenough, drawing on over two decades of experience studying the electronic and magnetic properties of transition-metal oxides at Lincoln Laboratory, Massachusetts, began focusing on oxides as cathodes after joining the University of Oxford in 1976. Kiochi Mizushima, who came on leave from Tokyo University as a visiting scientist to work with Goodenough, identified layered lithium cobalt oxide (LiCoO2) as a cathode in 1980 (Mizushima et al. 1980). As with TiS2, the direct metal-metal (Co-Co) interaction and the two-dimensional layered structure facilitate good electronic and lithium-ion conductivity. Akira Yoshino at Asahi Kasei Corporation in Japan worked on carbonaceous materials as anodes. He demonstrated the first rechargeable lithium-ion cell, employing LiCoO2 cathode and a carbon-based anode (petroleum coke) in 1985 (Yoshino et al. 1987). Following this, Sony Corporation commercialized lithium-ion technology in 1991 with LiCoO2 cathode and a coke anode. Later, in 1997, the industry transitioned to graphite anode as it displays a flatter discharge voltage (Figure 2). For their work, Whittingham, Goodenough, and Yoshino were awarded the Nobel Prize in Chemistry in 2019. Commercialized lithium-ion technology offered two critical advantages. First, replacing a sulfide cathode like TiS2 with an oxide cathode like LiCoO2 enabled a higher cell voltage of ~4 V versus <2.5 V with a sulfide cathode. Second, because the cathode already contained lithium in it, it eliminated the need to use lithium metal as an anode and allowed lithium-free anodes like graphite. Present-day lithium-ion technology is based on this “rocking chair” concept in which the lithium ions shuttle between the two electrodes during the charge–discharge process without involving metallic lithium. The Cathode Challenge About 75% of the cost of a lithium-ion battery comes from its materials, with roughly half of that attributed to the cathode, which relies on costly metals like cobalt and nickel (Li et al. 2020). The cathode also imposes key limits on energy density due to surface instability with liquid organic electrolytes, restricting the operating voltage to below 4.3 V, and a relatively low charge-storage capacity (<220 A h kg–¹) compared with ~370 A h kg–¹ for graphite anodes. In addition, oxide cathodes can release oxygen gas during overcharge, posing fire hazards and safety concerns. Following the discovery of layered LiCoO2, lithium manganese oxide (LiMn2O4) with a spinel structure was identified in 1983 by Michael Thackeray, who, while on leave from the Council for Scientific and Industrial Research in South Africa, worked with Goodenough as a visiting scientist at Oxford (Thackeray et al. 1983). The three-dimensional spinel framework, characterized by direct Mn–Mn interactions and 3D lithium-ion diffusion pathways, enabled faster charge–discharge rates and greater structural stability compared to the two-dimensional layered LiCoO2. LiMn2O4 also exhibited reduced oxygen release, improved safety, and a significant cost advantage since Mn is abundant and inexpensive relative to cobalt. However, dissolution of Mn²+ in the presence of trace protons in liquid electrolytes, followed by its migration to the anode, catalyzes electrolyte reduction on the graphite surface, increases cell impedance, and shortens cycle life. Consequently, despite its favorable cost and safety profile, LiMn2O4 could not be widely adopted due to the persistent issue of manganese dissolution. Since the commercialization of LiCoO2 in 1991, the prevailing trend over the past 35 years has been to progressively substitute cobalt with nickel and manganese to produce layered lithium nickel manganese cobalt oxide cathodes (LiCo1–2xMnxNixO2). This strategy is motivated by two factors: (1) both manganese and nickel are less expensive than cobalt, and (2) Ni³+ can be oxidized nearly to Ni4+, enabling higher charge-storage capacity compared with LiCoO2 (Chebiam et al. 2001). Consequently, cathode compositions evolved from LiCoO2 to LiNi1/3Mn1/3Co1/3O2 (NMC 111), LiNi0.6Mn0.2Co0.2O2 (NMC 622), and LiNi0.8Mn0.1Co0.1O2 (NMC 811), as well as LiNi0.8Co0.15Al0.05O2 (NCA) (Figure 2). Today, NMC 811 is widely regarded as the industry standard, while R&D efforts continue to push nickel content toward ~90% compositions that are beginning to see adoption. This compositional shift increases charge-storage capacity from ~150 A h kg–¹ for LiCoO2 to ~230 A h kg–¹ for LiNiO2, translating into higher energy densities. More recently, the Manthiram group demonstrated cobalt-free layered oxide cathodes, such as LiNi0.9Mn0.05Al0.05O2(NMA), which are now being manufactured by TexPower in Houston (Li et al. 2020). The Departure from Oxide Cathodes to Find a Pathway to Utilize Iron Iron is the least expensive transition metal and the fourth most abundant element in Earth’s crust. Its extraction and processing are low-cost and it is used in a variety of applications like steel. Despite its ubiquity, no way had been found to employ iron oxides as cathodes for lithium-ion batteries until recently. When Arumugam Manthiram arrived at Oxford in 1985 to work with Goodenough as a visiting scientist, cobalt- and manganese-based oxides (LiCoO2 and LiMn2O4) were known cathodes, but iron oxides were not. Drawing on his background in polyanion oxide chemistry from earlier work on lanthanide molybdates in India (Manthiram and Gopalakrishnan 1984), Manthiram first explored lithium insertion/extraction in the polyanion oxide iron molybdate (Fe2(MoO4)3) and subsequently in iron tungstate (Fe2(WO4)3) (Manthiram and Goodenough 1987). Both compounds exhibited a cell voltage of ~3 V with flat voltage profiles—significantly higher than that of Fe2O3 (<2.5 V)—despite all being governed by the Fe²+/Fe³+ redox couple. In 1986, Manthiram and Goodenough moved to The University of Texas at Austin (UT Austin), where Manthiram turned to iron sulfate (Fe2(SO4)3). Remarkably, it displayed an even higher cell voltage of ~3.6 V compared to Fe2(MoO4)3 and Fe2(WO4)3, again with the same structure and the same Fe²+/Fe³+ redox couple (Manthiram and Goodenough, 1989). The marked increase in cell voltage—by more than 1 V—when moving from a simple oxide such as Fe2O3 to polyanion oxides led Manthiram and Goodenough to recognize the inductive effect of counter cations (Mo6+, W6+, or S6+) on lowering the redox energy of Fe²+/³+ in polyanion oxides compared to that in an oxide (Fe2O3) and thereby raising the cell voltage. The more covalent Mo-O and W-O bonds compared to the Fe-O bonds weaken the Fe-O covalency through inductive effect and lower the Fe2+/3+ redox energy. The even more covalent S-O bond compared to the Mo-O and W-O bonds weakens the Fe-O covalence further, thereby lowering the Fe2+/3+ redox energy even much more (Manthiram 2020) and raising the cell voltage even further. Intrigued by these findings, a PhD student Geeta Ahuja then pursued with Manthiram and Goodenough transition-metal phosphates, which comprised part of her dissertation in 1991 (Ahuja 1991). Based on the above foundation, Goodenough with his students identified lithium iron phosphate (LiFePO4) with the olivine structure and a flat 3.5 V profile as a cathode in 1997 (Padhi et al. 1997) (Figure 2)—a decade later after Manthiram laid the groundwork with polyanion oxide pathway to employ iron in cathodes (Manthiram and Goodenough 1987, 1989). Collectively, the work of Manthiram and Goodenough in the late 1980s led to: establishing a pathway to employ iron—the least expensive metal with suitable properties—in lithium-ion batteries; opening the broad field of polyanion cathodes, including LiFePO4 (LFP), Li3V2(PO4)3 (LVP), Na3V2(PO4)3 (NVP), Na3V2(PO4)2F3 (NVPF), and others for both lithium-ion and sodium-ion batteries (Masquelier and Croguennec 2013); uncovering the inductive effect of counter-cations, which enabled higher cell voltages with more stable, lower-valent redox couples such as Fe²+/³+; and improving thermal stability, safety, and cycle life through the tightly bound oxygen in covalently bonded polyanion groups and the use of stable lower-valent redox couples. By 2024, about 40% of the lithium-ion battery market—worth $24 billion USD—was based on LFP cathodes. The abundance, low cost, and supply-chain advantages of iron, combined with the enhanced safety of polyanion oxide cathodes, are now driving industry adoption of LFP even further. Its market share is projected to rise well above 40%, particularly with the rapid growth of energy storage applications. However, despite the notable cost, supply-chain, and safety advantages of polyanion oxide cathodes, LFP has drawbacks. It is less dense than oxide cathodes such as NMC and is intrinsically a poor electronic and ionic conductor, which limits charge transport. An analogous polyanion oxide, LiMn1–xFexPO4 (LMFP) has recently attracted considerable interest. LMFP operates at a higher voltage, thereby offering higher energy density. The unique opportunity for three visiting researchers from Japan, South Africa, and India to work with Goodenough in the 1980 without overlapping led to the development of three distinct families of cathodes: layered oxides, spinel oxides, and polyanion oxides. Of these, layered oxides and polyanion oxides are now widely used in commercial lithium-ion batteries. Looking ahead, the cost and supply-chain benefits of iron and manganese, coupled with improved safety and longer cycle life, are expected to further expand the market share of LFP and LMFP cathodes in both Western countries and emerging markets such as India. Moreover, blending a cobalt-lean or cobalt-free layered NMC oxide with LMFP or LFP can reduce costs while minimizing energy density penalties, and at the same time improve safety and cycle life compared to using NMC alone (Lee et al. 2024). Moving Forward Cost, sustainability, and supply chain challenges are driving interest in alternative materials and battery chemistries. Examples include lithium-ion batteries based on more abundant metals, sodium-ion batteries (using sodium instead of lithium as the working ion), and lithium-sulfur or sodium-sulfur batteries, in which sulfur serves as the cathode without additional metals (Figure 2). However, most of these technologies remain at the R&D or prototype stage. With lithium-ion batteries, there is significant interest in replacing graphite with silicon anodes, as silicon is abundant and offers an order-of-magnitude higher charge-storage capacity. Unfortunately, large volume changes during cycling and aggressive surface reactivity with the liquid electrolyte severely shorten cycle life. As a compromise, commercial cells currently use silicon-graphite composite anodes containing <10% silicon (Figure 2). Renewed efforts are also underway to employ lithium metal anodes. For instance, the US Department of Energy–funded Battery500 Consortium is targeting an energy density of 500 W h kg–1 with lithium-metal anodes, compared to ~300 W h kg–1 in today’s state-of-the-art lithium-ion cells. In sodium-ion batteries, layered oxides can be synthesized largely with manganese and iron, minimizing expensive nickel and further lowering costs. Other promising cathodes include polyanion oxides such as NVP and NVPF, and Prussian blue analogs such as Na2FeFe(CN)6 and Na2FeMn(CN)6. While Prussian blue analogs are low cost, they suffer from toxic gas release and safety concerns, and both these and polyanion cathodes deliver lower energy density than layered oxides. Sodium-ion batteries also require hard carbon anodes (short-range order) instead of graphite (long-range order), lowering the cell voltage and complicating performance consistency. Overall, sodium-ion batteries typically exhibit lower energy density than lithium-ion batteries, and the current goal is to bring sodium-ion technology to parity with LFP-based lithium-ion cells. Sulfur provides notable benefits, including an order-of-magnitude higher capacity than oxide or polyanion cathodes, broad abundance, favorable supply-chain availability, and low cost as a petrochemical byproduct. Sodium–sulfur batteries, in particular, can be considered “mined-metal-free” since sodium is widely available in seawater. However, sulfur’s poor ionic and electronic conductivity necessitates large amounts of conductive carbon and liquid electrolyte, reducing the practical energy density. Moreover, during discharge, soluble higher-order polysulfides form and shuttle between the sulfur cathode and lithium- or sodium-metal anode, causing severe capacity fading. Thus, despite sulfur’s compelling cost and supply-chain advantages, fundamental scientific challenges must be overcome for commercialization. Electrolytes play a pivotal role in all of these battery chemistries. With the high negative electrochemical reduction potentials, anodes (graphite, silicon, lithium metal, hard carbon, sodium metal) and highly oxidizing cathodes (e.g., NMC) react with liquid electrolytes, forming solid-electrolyte interphase (SEI) and cathode-electrolyte interphase (CEI) layers. These interphases consume active lithium or sodium and impede ionic transport. Complex solvent and additive blends that are often found through trial-and-error are used to mitigate these effects. Challenges are especially severe with metal anodes due to repeated plating/stripping and new surface formation, as well as with silicon anodes because of volume expansion and nanoscale particle size. Flammable solvents further raise fire-safety risks. While the solid electrodes or electrolytes with fixed atomic positions in the crystal lattice are more straightforward to model and characterize, liquid electrolytes with multiple solvents and salts present far greater complexity. Solvation structures and the associated solvation/desolvation dynamics strongly influence SEI, CEI, and overall performance. Here, AI tools can accelerate the discovery of high-performance electrolyte formulations with balanced compatibility across electrodes. Cost, sustainability, and supply chain challenges are driving interest in alternative materials and battery chemistries. To address liquid-electrolyte limitations, all-solid-state batteries are being aggressively pursued with lithium- and sodium-metal anodes and both oxide and sulfur cathodes (Figure 2). Solid-state designs promise higher energy density by leveraging high-capacity metal anodes and eliminating the weight of liquid electrolytes, while also improving safety by removing flammable solvents. Yet, ion transport across solid–solid interfaces remains sluggish, and manufacturing large, defect-free separators with today’s ceramic oxide or sulfide electrolytes is difficult. Solid electrolytes can also form undesirable SEI or CEI layers, depending on material pairings. For large-scale grid storage employing batteries, cost and reliability become critical. In this context, redox flow batteries are being developed with aqueous and nonaqueous liquid electrodes (Figure 2). A key challenge with this technology is species crossover through conventional porous polymer separators. One strategy to mitigate crossover is to use solid-state electrolyte separators, which conduct ions through lattice sites without pores (Manthiram et al. 2017). This approach enables hybrid chemistries, for example coupling aqueous and nonaqueous, or acidic and basic electrodes within a single system. Empowering Accelerated Development with AI A vast amount of experimental and computational data accumulated over more than four decades with various battery chemistries and materials are available in the published literature, public databases, and in industry. As we march into the “fourth paradigm” (data-driven approach) of materials research driven by data, advances in AI are rapidly reshaping with respect to how researchers discover and optimize battery materials (Lombardo et al. 2022; Wang et al. 2024). The premise is whether one can use the wealth of data available with AI and machine learning (ML) to accelerate the developments with the desired battery performance metrics. AI/ML Basics AI broadly refers to computer systems that perform tasks usually requiring human intelligence; ML is a core subset of AI that uses algorithms or statistical models to learn patterns from data, enabling predictive modeling without explicit rule-based programming. ML algorithms mainly consist of unsupervised learning, supervised learning, and reinforcement learning. Supervised learning trains models on labeled data (i.e., the target or outcome variable is known) to predict outputs for new, unseen inputs. This method is widely used in applications like image recognition, speech recognition, spam detection, and predictive analytics. These models learn to map inputs to desired outputs by minimizing error on known examples. In battery research, supervised learning is applied to predict key materials properties based on structural, chemical, or thermodynamic descriptors. Common tasks include regression (e.g., predicting continuous values like ionic conductivity, voltage, or capacity) and classification (e.g., determining whether a material is stable within a voltage window or whether a certain formulation leads to dendrite formation). In contrast, unsupervised learning is used when the data lacks explicit labels. This method aims to detect underlying patterns, groupings, or structures within data without prior annotation. In the context of battery research, unsupervised learning can be applied to cluster materials based on structural or compositional similarity, reveal trends in electrolyte solvation structures, or reduce the dimensionality of complex simulation data. Self-supervised learning is a specialized form of unsupervised learning in which models generate pseudo-labels from unlabeled data, enabling them to capture meaningful patterns and feature representations without extensive manual annotation. This approach has become central in natural language processing (NLP) and computer vision, where models are pre-trained on large datasets by exploiting inherent data structures and then fine-tuned for specific applications. In materials engineering and battery research, self-supervised learning can be applied to molecular graphs, atomic environments, or spectral data to extract rich descriptors that are later fine-tuned for tasks such as property prediction or structure classification. Meanwhile, there is also semi-supervised learning, which combines supervised learning and unsupervised learning by using both labeled and unlabeled data to train the models. This method leverages a small set of labeled data along with a large amount of unlabeled data to improve learning efficiency. It is particularly useful in situations where obtaining a sufficient amount of labeled data can be expensive or time-consuming, but large amounts of unlabeled data are relatively easy to acquire. Reinforcement learning (RL) is an ML technique that is used to guide decision-making processes. It involves an autonomous agent that interacts with an environment and learns optimal strategies through trial and error in the absence of any guidance from a human user. In battery research, RL can be used to optimize formulations, cycling protocols, or synthesis conditions with minimal experiments. When coupled with automated experimentation in closed-loop systems, they enable self-driving labs that rapidly converge on optimal materials candidates, significantly reducing development time and experimental costs. As the traditional ML techniques reach their limitations in handling large volumes of high-dimensional, unstructured data, deep learning (DL) has emerged as a powerful alternative to tackle the challenges by utilizing multilayered neural networks. Different forms of DL have been used in many of the AI applications in our lives today, such as ChatGPT and self-driving cars. In materials engineering, DL enables the direct use of raw or minimally processed data, such as atomic configurations, spectral data, and crystallographic data to model target properties. These models can learn hierarchical and spatial features automatically, making them especially powerful for tasks like capturing complex structure-property relationships, representing compositional effects, or even generating new candidate materials in a multi-dimensional chemical space. Moreover, concepts like active learning, which emphasizes efficiency by enabling AI systems to identify the most informative data points for labeling or experimentation, and transfer learning, which allows models trained on one type of material system to be adapted to another, are also promising ML techniques that can be employed in battery materials research. AI-Assisted Data Acquisition At the heart of all the ML approaches lies data: AI and ML flourish only when presented with abundant, diverse, high-quality datasets to learn from (Figure 3) and the performance of AI models is tightly coupled to the diversity, relevance, and quality of the data they are trained on. AI learns by example, so the effectiveness of a model depends on how well the training data represents the chemical and structural landscape of interest. Therefore, before training AI models for battery-related research, they need to be provided with rich and accessible datasets that accurately reflect the complex behavior of battery materials. This includes experimental data from electrochemical tests, physical characterization, and spectroscopy, computational data from first-principles calculations, and molecular dynamics simulations. A straightforward approach to data acquisition is mining published literature and patents, which contain decades of experimental and computational results. While searchable databases provide direct access, they may lack the specific data types or scale needed for battery researchers to train ML models. In such cases, AI can help through automated literature mining (Kononova et al. 2021). Modern NLP tools, including large language models (LLMs), can extract structured data from unstructured text, allowing researchers to collect property measurements, synthesis parameters, and performance metrics from thousands of scientific documents (Miret and Krishnan 2025). Beyond mining existing resources, another way to acquire battery-relevant data is by generating large-scale datasets through high-throughput computational or experimental methods augmented by AI. On the computational side, ML-based interatomic potentials and surrogate models can deliver accuracy comparable to costly density functional theory (DFT) calculations, but at a fraction of the time. These models have been applied to predict electronic structures, density, viscosity, ionic conductivity, and other physicochemical properties (Gong et al. 2025). The resulting simulated datasets can then serve as inputs for downstream ML pipelines or as validation sets in materials screening workflows. For example, NVIDIA’s collaboration with SES AI employed GPU-accelerated ML to solve structures and compute electronic properties for more than 100 million candidate electrolyte molecules—including highest occupied molecular orbital (HOMO) and lowest unoccupied molecular orbital (LUMO) levels as well as electrostatic potentials—effectively mapping the “molecular universe” of battery chemistry with unprecedented efficiency (Xu et al. 2025). Additionally, high-throughput experimentation can generate experimental data orders of magnitude faster than manual approaches. The integration of robotics and AI into laboratory automation has given rise to self-driving laboratories that autonomously plan, execute, and analyze experiments. For instance, researchers have developed robotic systems that move between stations, mixing and testing compounds under AI guidance. One of the earliest demonstrations came from the University of Liverpool, where a “mobile robotic chemist” autonomously explored a photocatalyst formulation space, conducting 688 experiments in just eight days (Burger et al. 2020). Battery researchers have likewise been early adopters. At Carnegie Mellon University (CMU), scientists developed the battery-focused test stand “Otto,” which screened 140 electrolyte formulations in 40 hours (Dave et al. 2020). Such automated systems are often integrated with active learning or Bayesian optimization to create closed-loop frameworks that continuously refine experimental strategies based on feedback, accelerating the discovery of promising new materials. Predicting and Optimizing Battery Materials With carefully designed algorithms and proper training on high-quality datasets, AI methods can be applied to specific battery materials and properties (Lombardo et al. 2022) (Figure 4). In cathode discovery, models are used to predict intrinsic properties, such as crystal structure, density, and conductivity, as well as performance metrics, such as capacity, cycle life, redox energy, thermal stability, and rate capability. AI tools can screen vast compositional spaces, identify promising dopant elements, and estimate electrochemical performance based on structure and chemical composition. A recent example is DRXNet, a DL model developed by Gerbrand Ceder’s group (Zhong et al. 2024). DRXNet predicts accessible capacities and voltage profiles for new Li–Mn–O–F compositions as well as for high-entropy disordered rock salt (DRX) systems containing diverse metal species. By serving as a universal surrogate, it enables rapid screening of thousands of hypothetical DRX cathodes, thereby accelerating the discovery of new electrode materials. AI is also increasingly applied to electrolyte design. For liquid electrolytes, ML models can predict critical properties such as ionic conductivity, interfacial stability, and electrochemical stability windows (oxidative/reductive limits) (Kumar et al. 2025). ML and surrogate models have also been used to study solvation structures and intermolecular interactions (Gao et al. 2023). Solid-state electrolytes, another major focus area, similarly benefit from ML predictions of ionic conductivity, structural stability, and electrochemical compatibility. Recent advances in DL have enabled models that operate directly on unstructured inputs such as molecular graphs or atomistic geometries. These methods are particularly powerful for capturing subtle structure–property relationships across high-dimensional chemical spaces, reducing reliance on expert intuition and improving predictive performance (Zhang et al. 2024). In parallel, state-of-the-art generative ML approaches are being developed to propose entirely new electrolyte molecules or mixtures, for both liquid and solid-state systems. By learning patterns from existing electrolyte databases and sampling new molecular graphs with desired properties, these models can effectively “invent” novel chemistries (Wang 2025; Yang et al. 2025). In essence, AI is transforming battery R&D into a data-driven optimization problem: given design targets, it can navigate multidimensional chemical space to suggest top candidates and experimental plans. AI-Driven Battery Research Efforts Several high-profile projects exemplify the AI-driven paradigm, particularly in demonstrating the power of closed-loop operations where AI systems iteratively generate hypotheses, guide experiments, and refine their models with newly acquired data. These autonomous cycles of prediction, testing, and learning are transforming materials discovery from a manual, intuition-driven pursuit into a scalable, data-centric process. A few illustrative examples are highlighted below. CMU’s Clio A team from CMU demonstrated an autonomous workflow combining active learning with high-throughput experimentation to optimize liquid electrolytes for lithium batteries (Dave et al. 2022). The platform iteratively tested candidate mixtures in coin cells, using Bayesian optimization to prioritize formulations that improved ionic conductivity and electrochemical stability. Over a series of closed-loop cycles, the system converged on novel electrolyte compositions that outperformed benchmark formulations, showcasing the power of AI-guided experimentation to accelerate materials optimization. Google DeepMind’s GNoME (graph networks for materials exploration) In 2023, DeepMind used deep graph neural networks to explore inorganic crystal space, predicting 2.2 million new crystalline compounds, of which about 380,000 were calculated to be thermodynamically stable (Merchant et al. 2023). Notably for battery research, GNoME identified 528 potential lithium-ion conductors, any of which could serve as novel solid electrolytes or cathode components. The model was trained on existing materials databases (e.g., the Materials Project) and employed active learning loops—predicting candidates and then verifying them with DFT—to achieve this unprecedented scale. DeepMind has released its 380,000 most stable predictions publicly. While some concerns remain regarding the true novelty of these materials (Cheetham and Seshadri 2024), collaborators at Lawrence Berkeley National Laboratory have already begun using these predictions to guide autonomous synthesis efforts, as discussed in the next example. Autonomous Laboratories (Robotic Chemists, A-Lab) Beyond predictive modeling, the emergence of fully automated laboratories marks a significant advance in closed-loop experimentation for battery materials. A-Lab, developed at Berkeley Lab, exemplifies this paradigm by integrating robotic synthesis, automated characterization, and AI-driven decision making into a self-driving platform for materials discovery (Szymanski et al. 2023). The system operates by autonomously selecting candidate materials, synthesizing and characterizing them, and using ML to analyze the resulting data and propose the next set of experiments—all with minimal human intervention. In their demonstration, the A-Lab team synthesized and characterized more than 41 previously unreported inorganic compounds, showcasing the potential of autonomous labs to dramatically accelerate discovery and ease experimental bottlenecks in battery materials innovation. Looking ahead, these platforms could be extended beyond synthesizability to investigate factors such as microstructure and materials performance. Microsoft Azure and Pacific Northwest National Laboratory Collaboration In early 2024, Microsoft and the Pacific Northwest National Laboratory showcased how AI and cloud-scale computing can compress years of discovery into months (Chen et al. 2024; Xu et al. 2024). Using Azure Quantum Elements, AI models narrowed a pool of ~32 million inorganic formulas to about 500,000 predicted to be thermodynamically stable. From this set, ~18 candidates were selected (in roughly 80 computing hours) for synthesis and testing in a battery context. One of these proved to be a solid-state electrolyte that synergistically combined lithium and sodium ions. This material reduced lithium content requirements by ~70% while maintaining ionic conductivity, challenging prior assumptions about the stability of mixed-ion systems. Future Outlook: Challenges and Opportunities The AI revolution in battery research is still in its early stages, and progress on multiple fronts is needed to realize its full potential. First, experimental infrastructure must expand. Automated laboratories with broader capabilities, including high-throughput synthesis, characterization, and electrochemical testing, are essential for generating reliable data at scale. In the coming years, AI could become as indispensable to battery research as electrochemistry itself. Second, data quality and accessibility must be improved. Beyond existing public databases such as Materials Project, AFLOW, and OQMD, training powerful ML models will require large, high-quality, open-source datasets of battery materials spanning diverse chemistries and conditions. Equally important are standardized protocols for reporting experimental and computational results, including metadata such as synthesis or fabrication details, test conditions, and computational parameters. Such standardization would enhance data interoperability and reusability in line with the FAIR (Findable, Accessible, Interoperable, and Reusable) principles, which are critical for data-driven approaches. At present, much of the published literature lacks accurate metadata and consistent evaluation methods. Addressing these gaps will require coordinated community efforts such as the proposed “Battery Data Genome” initiative (Ward et al. 2022) along with incentives for companies and labs to share data. Third, new AI tools and platforms may need to be tailored specifically for battery research. Advances in accuracy and the development of multiscale models will be critical to link atomic-scale predictions with real-world device performance, bridging gaps between atomistic ML models, continuum battery models, and circuit-level simulations. Equally important is usability: accessible ML frameworks, cloud-based platforms, and educational resources can lower barriers to adoption and enable broader use of AI among battery scientists. In the coming years, AI could become as indispensable to battery research as electrochemistry itself. By combining human ingenuity with machine intelligence, the field can more rapidly identify breakthrough materials, whether high-capacity cathodes, safer electrolytes and electrodes, or durable solid-state systems. Ultimately, an ecosystem of standardized data, interpretable models, and autonomous labs could transform battery R&D from an artisanal trial-and-error practice into a data- and AI-driven engineering discipline. Seamless collaboration and communication between experienced experimentalists and computational experts will be essential, ensuring the human expertise needed to guide and accelerate discoveries in energy storage. Summary Half a century of concerted basic science and engineering research has led to today’s lithium-ion battery technology. About 50% of materials cost is from the cathode, and basic science research in the 1980s offered pathways to lower the cost and ease supply chain—from cobalt oxide to manganese oxide to iron polyanion oxide cathodes. As the battery market expands to large-scale applications such as EVs and grid storage, cost and supply-chain challenges are poised to become increasingly critical. Efforts must focus on eliminating scarce critical metals such as cobalt and nickel, while advancing battery chemistries based on earth-abundant elements like iron, manganese, sulfur, and sodium, as well as organic materials. Replacing liquid electrolytes with solid-state alternatives offers the potential to improve safety and increase energy density. With the vast troves of experimental data available in the literature and industry, AI can help accelerate the discovery of cost-effective, supply-chain-resilient chemistries and materials. Appropriately combining machine intelligence with human ingenuity could enable rapid, transformative advances in energy storage. Acknowledgments The authors acknowledge support by the US Department of Energy, Office of Basic Energy Sciences, Division of Materials Science and Engineering under award number DE-SC0005397. References Ahuja G. 1991. An Investigation of Some Lithium Insertion Compounds. PhD Dissertation, Chapter 5, pp. 92–114, The University of Texas at Austin. Baker N. 2024. Unlocking a new era for scientific discovery with AI: How Microsoft’s AI screened over 32 million candidates to find a better battery. January 9. Online at https://azure.microsoft.com/en-us/blog/quantum/2024/01/ 09/unlocking-a-new-era-for-scientific-discovery-with-ai- how-microsofts-ai-screened-over-32-million-candidates- to-find-a-better-battery. Burger B, Maffettone PM, Gusev VV, Aitchison CM, Bai Y, Wang X, Li X, Alston BM, Li B, Clowes R, and 4 others. 2020. A mobile robotic chemist. Nature 583:237–41. Chebiam RV, Kannan AM, Prado F, Manthiram A. 2001. Comparison of the chemical stability of the high energy density cathodes of lithium-ion batteries. Electrochemistry Communications 3(11):624–27. Cheetham AK, Seshadri R. 2024. Artificial intelligence driving materials discovery? Perspective on the article: Scaling deep learning for materials discovery. Chemistry of Materials 36(8):3490–95. Chen C, Nguyen DT, Lee SJ, Baker NA, Karakoti AS, Lauw L, Owen C, Mueller KT, Bilodeau BA, Murugesan V, Toyer M. 2024. Accelerating computational materials discovery with machine learning and cloud high-performance computing: From large-scale screening to experimental validation. Journal of the American Chemical Society 146(29):20009–18. Dave A, Mitchell J, Kandasamy K, Wang H, Burke S, Paria B, Póczos B, Whitacre J, Viswanathan V. 2020. Autonomous discovery of battery electrolytes with robotic experimentation and machine learning. Cell Reports Physical Science 1(12):100264. Dave A, Mitchell J, Burke S, Lin H, Whitacre J, Viswanathan V. 2022. Autonomous optimization of non-aqueous Li-ion battery electrolytes via robotic experimentation and machine learning coupling. Nature Communications 13:5454. Gao Y, Yao N, Chen X, Yu L, Zhang R, Zhang Q. 2023. Data-driven insight into the reductive stability of ion–solvent complexes in lithium battery electrolytes. Journal of the American Chemical Society 145(43):23764–70. Gong S, Zhang Y, Mu Z, Pu Z, Wang H, Han X, Yu Z, Chen M, Zheng T, Wang Z, and 7 others. 2025. A predictive machine learning force-field framework for liquid electrolyte development. Nature Machine Intelligence 7:543–52. Kononova O, He T, Huo H, Trewartha A, Olivetti EA, Ceder G. 2021. Opportunities and challenges of text mining in materials research. iScience 24(3):102155. Kumar R, Vu MC, Ma P, Amanchukwu CV. 2025. Electrolytomics: A unified big data approach for electrolyte design and discovery. Chemistry of Materials 37(8):2720–34. Lee S, Scanlan K, Reed S, Manthiram A. 2025. Cost-effective layered oxide—olivine blend cathodes for high-rate pulse power lithium-ion batteries. Advanced Energy Materials 15(5):2403002. Li W, Erickson E, Manthiram A. 2020a. High-nickel layered oxide cathodes for lithium-based automotive batteries. Nature Energy 5(1):26–34. Li W, Lee S, Manthiram A. 2020b. High-nickel NMA: A cobalt-free alternative to NMC and NCA cathodes for lithium-ion batteries. Advanced Materials 32(33):2002718. Lombardo T, Duquesnoy M, El-Bouysidy H, Årén F, Gallo-Bueno A, Jørgensen PB, Bhowmik A, Demortière A, Ayerbe E, Alcaide F, and 7 others. 2022. Artificial intelligence applied to battery research: Hype or reality? Chemical Reviews 122(12):10899–969. Manthiram A. 2020. A reflection on lithium-ion battery cathode chemistry. Nature Communications 11(1):1550. Manthiram A, Goodenough JB. 1987. Lithium insertion into Fe2(MO4)3 frameworks: Comparison of M = W with M = Mo. Journal of Solid State Chemistry 71(2):349–60. Manthiram A, Goodenough JB. 1989. Lithium insertion into Fe2(SO4)3 framework. Journal of Power Sources 26(3–4):403–8. Manthiram A, Gopalakrishnan J. 1984. Lower-valence molybdenum oxides: Crystal chemistry and electronics properties. Reviews in Inorganic Chemistry 6:1–68. Manthiram A, Yu X, Wang S. 2017. Lithium battery chemistries enabled by solid-state electrolytes. Nature Reviews Materials 2:16103. Masquelier C, Croguennec L. 2013. Polyanionic (phosphates, silicates, sulfates) frameworks as electrode materials for rechargeable Li (or Na) batteries. Chemical Reviews 113(8):6552–91. Merchant A, Batzner S, Schoenholz SS, Aykol M, Cheon G, Cubuk ED. 2023. Scaling deep learning for materials discovery. Nature 624:80–5. Miret S, Krishnan NMA. 2025. Enabling large language models for real-world materials discovery. Nature Machine Intelligence 7:991–8. Mizushima K, Jones PC, Wiseman PJ, Goodenough JB. 1980. LixCoO2 (0 < x < 1): A new cathode material for batteries of high energy density. Materials Research Bulletin 15(6):783–98. Padhi AK, Nanjundaswamy KS, Goodenough JB. 1997. Phospho-olivines as positive electrode materials for rechargeable lithium batteries. Journal of the Electrochemical Society 144:1188–94. Szymanski NJ, Rendy B, Fei Y, Kumar RE, He T, Milsted D, McDermott MJ, Gallant M, Cubuk ED, Merchant A, and 6 others. 2023. An autonomous laboratory for the accelerated synthesis of novel materials. Nature 624:86–91. Thackeray MM, David WIF, Bruce PG, Goodenough JB. 1983. Lithium insertion into manganese spinels. Materials Research Bulletin 18(4):461–72. The Research Insights. 2025. Lithium ion battery market. Online at https://www.theresearchinsights.com/reports/lithium-ion- battery-market-998. Wang Y. 2025. Application-oriented design of machine learning paradigms for battery science. npj Computational Materials 11:89. Wang Z, Chen A, Tao K, Han Y, Li J. 2024. MatGPT: A vane of materials informatics from past, present, to future. Advanced Materials 36(6):2306733. Ward L, Babinec S, Dufek EJ, Howey DA, Viswanathan V, Aykol M, Beck DAC, Blaiszik B, Chen BR, Crabtree G, and 18 others. 2022. Principles of the battery data genome. Joule 6(10):2253–71. Whittingham, MS. 1976. Electrical energy storage and intercalation chemistry. Science 192(4244):1126–7. Xu K, Hannah D, Zhang Y, Hu Q. 2025. Spotlight: Accelerating the discovery of new battery materials with SES AI’s molecular universe. May 8. Online at https://developer.nvidia.com/blog/spotlight-accelerating- the-discovery-of-new-battery-materials-with-ses-ais- molecula r-universe. Yang Z, Wu Y, Han X, Zhang Z, Lai H, Mu Z, Zheng T, Liu S, Pu Z, Wang Z, and 3 others. 2025. A unified predictive and generative solution for liquid electrolyte formulation. arXiv. Online at https://doi.org/10.48550/arXiv.2504.18728. Yoshino A, Sanechika K, Nakajima T. 1987. Secondary battery. US Patent 4668595 and Japan Patent 1989293. Zhang H, Lai T, Chen J, Manthiram A, Rondinelli JM, Chen W. 2024. Learning molecular mixture property using chemistry-aware graph neural network. PRX Energy 3(2):023006. Zhong P, Deng B, He T, Lun Z, Ceder G. 2024. Deep learning of experimental electrochemistry for battery cathodes across diverse compositions. Joule 8(6):1837–54. About the Author:Arumugam Manthiram is George T. and Gladys H. Abell Endowed Chair of Engineering at The University of Texas at Austin. Tianxing Lai is graduate research assistant in the Manthiram Lab at The University of Texas at Austin