AI for Discovering New Materials

The Materials Discovery Challenge

Every technology we use depends on materials with specific properties. Batteries need materials that store and release lithium ions efficiently. Solar cells need materials that absorb sunlight and convert it to electricity. Airplane components need materials that are both strong and light. Finding the right material for a given application has historically been a slow, intuition-driven process. Thomas Edison tested over 3,000 materials for lightbulb filaments. Modern materials science is more systematic, but the fundamental challenge remains: the space of possible material compositions is vast, and testing each one experimentally is expensive and slow.

Consider the search for a new battery cathode material. A typical cathode contains three to five elements in specific ratios, with a specific crystal structure. The number of possible compositions with just the elements in the periodic table is enormous. Each candidate requires synthesis (days to weeks), characterization (days), and electrochemical testing (weeks to months). A research group might test 50 to 100 candidates per year. At that rate, finding the optimal composition by brute force would take centuries.

AI compresses this search by learning the relationship between composition, structure, and properties from existing data, then predicting which untested compositions are most likely to have the desired properties. Instead of testing 100 random candidates, a research group guided by AI tests the 100 most promising candidates, finding better materials faster.

Property Prediction from Composition and Structure

The foundational AI task in materials science is predicting material properties from a description of the material. Given the chemical formula and crystal structure, what will the material's bandgap be? Its thermal conductivity? Its mechanical strength? Its stability at high temperatures? Machine learning models trained on databases of experimentally measured properties can answer these questions with useful accuracy for many property types.

The Materials Project, a public database containing computed properties for over 150,000 inorganic materials, provides the training data for many AI property prediction models. Each entry includes the crystal structure, computed formation energy, electronic band structure, elastic properties, and other quantities calculated using density functional theory (DFT), a physics-based simulation method that is accurate but computationally expensive. AI models trained on this data can approximate DFT accuracy at a tiny fraction of the computational cost, enabling screening of millions of candidates.

Graph neural networks (GNNs) have emerged as the dominant architecture for materials property prediction. They represent the crystal structure as a graph, with atoms as nodes and bonds as edges, and learn to propagate information along the graph to predict properties. CGCNN (Crystal Graph Convolutional Neural Network), MEGNet, and M3GNet are widely used GNN models that predict formation energy, bandgap, and other properties with mean absolute errors comparable to the uncertainty in the DFT calculations they were trained on.

The accuracy of property prediction depends on the property and the data available. Properties that are determined primarily by composition (like formation energy) can be predicted with high accuracy. Properties that depend sensitively on defects, grain boundaries, or processing conditions (like fracture toughness or fatigue life) are much harder to predict from composition alone because these features are not captured in the idealized crystal structure description. For these properties, models that incorporate processing information and microstructural features are needed.

GNoME and Large-Scale Stability Prediction

Google DeepMind's GNoME (Graph Networks for Materials Exploration) system represents the largest AI materials discovery effort to date. GNoME used graph neural networks to predict the thermodynamic stability of 2.2 million new inorganic materials, 380,000 of which were predicted to be stable with high confidence. Independent experimental groups subsequently synthesized and confirmed over 700 of these predictions, validating the approach.

The significance of GNoME is primarily in scale. Before GNoME, about 48,000 stable inorganic materials were known. GNoME expanded this to potentially over 400,000, a tenfold increase that opens up new compositional spaces for every application. Among the predicted stable materials are potential new superconductors, thermoelectrics, battery materials, and catalysts. The predictions are publicly available, enabling researchers worldwide to explore this expanded materials landscape.

GNoME used an active learning loop: the model predicted stability for millions of candidates, the most uncertain predictions were verified with DFT calculations, and the verified results were added to the training set to improve the model. This iterative process progressively improved the model's accuracy while building a comprehensive database of computed material properties.

Generative Design of New Materials

Beyond predicting properties of known compositions, generative AI models can propose entirely new materials optimized for specific applications. These models learn the distribution of stable, useful materials from training data, then sample from that distribution to generate new candidates with desired properties.

Variational autoencoders (VAEs) and diffusion models are the most common generative architectures for materials. A VAE learns a compact representation of the space of materials, where similar materials are close together. You can then navigate this space toward regions with desired properties and decode new material compositions from those regions. Diffusion models generate crystal structures atom by atom, constrained to produce physically plausible arrangements.

A practical example: researchers at MIT used a generative model to design new solid-state electrolytes for lithium batteries. The model was trained on known lithium-conducting materials and asked to generate new compositions with high lithium conductivity and chemical stability. Of the 50 top-ranked generated materials, 5 were synthesized and tested, and 3 showed lithium conductivity superior to the best previously known solid electrolytes. The total time from model training to experimental validation was four months, compared to the years a traditional search would require.

High-Throughput Screening and Autonomous Labs

AI-guided high-throughput screening combines computational prediction with automated experimental testing. The workflow starts with AI predicting the most promising candidates from a large pool. These candidates are synthesized using automated or robotic platforms that can produce dozens of samples per day. Automated characterization measures the key properties. The results feed back into the AI model, which updates its predictions and suggests the next batch of candidates.

The A-Lab at Lawrence Berkeley National Laboratory demonstrated this concept by autonomously synthesizing and characterizing 41 new materials in 17 days with no human intervention. A robotic system weighed precursors, mixed them, heated them in a furnace, and analyzed the products with X-ray diffraction, all guided by an AI system that decided what to synthesize based on accumulated results. This level of automation transforms materials discovery from a years-long human-driven process into a weeks-long machine-driven one.

Autonomous labs are not yet common, but the components (robotic synthesis, automated characterization, AI decision-making) are available and becoming more accessible. Cloud-based platforms allow researchers to submit synthesis requests to remote automated labs, democratizing access to high-throughput experimentation. This trend is likely to reshape materials science over the next decade, making discovery faster and more systematic.

Applications by Sector

Energy storage. AI has identified new cathode materials for lithium-ion batteries with higher energy density and lower cobalt content, addressing both performance and supply chain concerns. Solid-state electrolytes predicted by AI promise safer batteries that do not catch fire. Beyond lithium, AI explores sodium-ion and potassium-ion battery materials as alternatives for grid-scale storage where cost matters more than energy density.

Solar energy. Perovskite solar cells, which can be made cheaply by printing rather than the expensive crystal growth required for silicon, have composition spaces that AI navigates effectively. AI models predict the stability of different perovskite compositions, addressing the key limitation that has prevented commercialization. Models also optimize tandem solar cell architectures, stacking different absorber materials to capture a wider range of the solar spectrum.

Catalysis. AI identifies new catalysts for industrial chemical reactions, hydrogen production, and CO2 conversion. Catalyst design is particularly well-suited to AI because the relationship between surface structure and catalytic activity is complex and non-linear, exactly the type of problem where machine learning outperforms human intuition. AI-discovered catalysts for ammonia synthesis, a process that consumes 1 to 2% of global energy, promise to reduce energy requirements significantly.

Structural materials. High-entropy alloys, which mix five or more elements in roughly equal proportions, exhibit unexpected combinations of strength, ductility, and corrosion resistance that classical metallurgy cannot predict. AI maps the vast compositional space of high-entropy alloys efficiently, identifying promising compositions for aerospace, biomedical implants, and extreme-environment applications.