AI Propels Rapid Discovery of Next-Generation Battery Materials

Artificial intelligence (AI) is accelerating the discovery of novel energy-storage materials, offering alternatives to traditional lithium-ion batteries. By bypassing decades of empirical trial-and-error, this computational approach reduces research timelines to mere months or days.

The Imperative for Lithium-Ion Alternatives

Lithium-ion batteries power everything from consumer electronics to electric vehicles (EVs) and grid-scale battery energy storage systems (BESS). However, surging global demand, supply-chain constraints, geopolitical friction, and the ecological footprint of lithium mining—such as high water consumption and chemical runoff—necessitate sustainable alternatives.

Safety is another critical driver. Volatile liquid electrolytes in traditional lithium-ion cells pose flammability and leakage risks; an internal separator failure can trigger catastrophic thermal runaway. This hazard drives research towards solid-state electrolytes, which replace liquids with stable solid compounds to deliver higher energy densities and inherent thermal stability.

How AI is Revolutionising Materials Discovery

Traditional materials research relies on slow, expensive empirical testing. AI, machine learning (ML), and deep learning transform this by analysing vast datasets of crystal structures, identifying complex quantum-chemical patterns, and predicting material properties. This accelerates the development lifecycle, from predicting atomic performance to optimising synthesis.

Expedited Screening and Prediction

AI can rapidly screen a vast configuration space of potential materials. Collaborative work by Microsoft and the Pacific Northwest National Laboratory (PNNL) used AI and cloud-based high-performance computing (HPC) to sift through 32 million theoretical inorganic materials, narrowing them down to 18 promising candidates in just 80 hours. This screening phase would have taken over two decades using conventional laboratory methods.

To evaluate solid-state electrolytes, AI models are trained to predict properties such as ionic conductivity, electrical conductivity, and chemical stability. Historically, calculating these properties required computationally expensive Density Functional Theory (DFT) and molecular dynamics simulations to model the transport of ions through a crystal lattice. AI surrogate models bypass this bottleneck by predicting the ionic conductivity ($\sigma$) of candidate structures using the Arrhenius relationship:

$$\sigma =\frac{{\sigma }_0}{T}\exp \left(-\frac{E_a}{k_{B}T}\right)$$

Where:

• $\sigma$ is the ionic conductivity.
• ${\sigma }_0$ is the pre-exponential factor, representing the frequency of ionic hopping attempts and lattice vibrations.
• $E_a$ is the activation energy barrier for ion migration through the crystallographic pathways.
• $k_B$ is the Boltzmann constant.
• $T$ is the absolute temperature.

By predicting $E_a$ and $\sigma$ in milliseconds rather than the hours required for conventional DFT calculations, machine learning models allow researchers to discard millions of low-performance compounds instantly.

Unconventional Material Combinations

AI can step outside the boundaries of conventional scientific intuition, proposing unique material combinations that human researchers might overlook. In the Microsoft-PNNL project, the AI identified a stable structure that mixes sodium and lithium ions within a single lattice.

Historically, material scientists avoided mixing these two alkali metals because their differing ionic radii—sodium ($N{a}^{+}$) at 1.02 Å and lithium ($L{i}^{+}$) at 0.76 Å—typically cause lattice strain, phase separation, and high energy barriers for ion movement. The AI, however, identified a stable quaternary crystal structure where sodium and lithium ions co-exist, leveraging the sodium pathways to stabilise the structural framework while allowing lithium ions to migrate with reduced activation energy.

Optimising Synthesis and Performance

Beyond theoretical discovery, AI is employed to optimise material synthesis and processing conditions, predict cell lifetime, and model degradation rates. Synthesis is often the ultimate bottleneck; a material may perform exceptionally well in simulations, but finding the exact thermodynamic pathway to synthesise it in a laboratory can take years.

AI models assist by predicting synthesisability scores and mapping out phase diagrams. Translating these AI-discovered compounds into physical devices relies heavily on advanced materials and interface engineering. Process equipment manufacturers like Yield Engineering Systems (YES) provide the vacuum curing, chemical vapour deposition (CVD), and precision cleaning tools required to prepare substrates and modify interfaces at the nanoscale. Incorporating these processing parameters into machine learning models improves property prediction accuracy, reducing activation energy prediction errors to a mean absolute error of less than $0.05\text{eV}$ compared to DFT, minimising interfacial resistance at solid-state boundaries to below $15\Omega {\text{cm}}^2$, and optimising crystallographic growth parameters within manufacturing tolerances of $\pm 2\%$.

Breakthroughs in Lithium-Reduced Batteries

The N2116 Solid-State Electrolyte

In a major breakthrough, the Microsoft and PNNL team identified a new solid-state electrolyte, named N2116, which has the potential to reduce lithium use by up to 70%. This material is a sodium-lithium-yttrium-chloride halide compound. Solid halide electrolytes are highly valued because they offer excellent electrochemical stability at high operating voltages and are softer than oxide-based alternatives, allowing for better interfacial contact with the electrodes.

The transition from the initial theoretical prediction to a working battery prototype with N2116 took a mere nine months. This demonstration showcases the unprecedented speed and accuracy that AI integration brings to the research pipeline.

Multivalent-Ion Battery Materials

Researchers at the New Jersey Institute of Technology (NJIT), led by Professor Dibakar Datta, have leveraged generative AI to discover new porous materials for multivalent-ion batteries. Unlike monovalent lithium-ion batteries, multivalent batteries utilise more abundant elements like magnesium, calcium, aluminium, and zinc. Because these ions carry multiple positive charges (such as $M{g}^{2+}$ or $A{l}^{3+}$), they transfer multiple electrons per ion during cycling, potentially offering significantly higher volumetric energy densities than standard lithium-ion systems.

The NJIT team's dual-AI approach combined a Crystal Diffusion Variational Autoencoder (CDVAE) with a Large Language Model (LLM). The CDVAE acts as a generative model that treats atoms as point clouds in a periodic space, diffusing them into stable crystallographic structures. Meanwhile, the LLM parses scientific literature to filter these candidates based on historical synthesis feasibility. This workflow successfully identified five novel porous transition metal oxide structures ideal for accommodating larger multivalent ions, bypassing traditional structural relaxation bottlenecks.

Solid Electrolyte Advancements

Stanford University researchers experimentally validated a chemical compound identified by AI as a promising solid electrolyte, known as LBS (${\text{Li}}_8{\text{B}}_{10}{\text{S}}_{19}$). This lithium borosulphide glass-ceramic demonstrates high electrochemical stability and can withstand high electrical loads without breaking down or forming dendrites, addressing a key limitation of existing solid-state interfaces.

By training machine learning classifiers on experimental data alongside computational databases, the Stanford team identified the specific structural phases of LBS that yield optimal ionic conductivity, allowing them to target their synthesis efforts with high precision.

Challenges and Future Outlook

Despite these rapid advancements, challenges remain in fully integrating AI into materials science. A major obstacle is the need for high-quality, standardised, and comprehensive datasets. Materials science data is frequently fragmented, inconsistent, and scattered across disparate publications. Crucially, there is a lack of negative data; journals rarely publish failed synthesis attempts, which introduces a severe positive bias that can mislead AI training algorithms.

The "black-box" nature of some deep learning models also raises concerns regarding interpretability. Without physical explainability, researchers cannot easily determine why a model predicts a specific material will succeed, necessitating careful verification. Consequently, while AI accelerates the discovery funnel, human expertise and rigorous experimental validation remain vital for testing structural stability, safety profiling, and scaling discoveries to commercial viability.

The future of AI in materials science involves real-time closed-loop analysis during active experimentation, and the application of transfer learning to export knowledge from well-characterised material systems to entirely unmapped chemical spaces. The long-term goal is for AI to evolve from a "co-pilot" assisting human researchers to an autonomous adviser capable of recommending synthesis pathways.

These self-driving laboratories—often called "A-Labs"—combine AI decision-making with robotic synthesis arms to generate, synthesise, and test material candidates in a continuous loop. Companies like Citrine Informatics and open-source initiatives like the Materials Project are at the forefront of this evolution, applying AI to predict material properties, construct rich datasets, and pioneer the next generation of energy storage solutions.