AI in materials and computational chemistry

How AI is used

Materials and computational chemistry ask what a substance is made of, how its atoms are arranged, and what properties follow. Answering that by physics alone is slow, so models are trained to shortcut it. Some learn the forces between atoms from quantum-mechanical calculations, then stand in for those calculations in simulations of much larger structures. Some read composition or crystal structure and predict a property: which crystal phase an alloy will form, how much gas a porous framework will soak up. Others work on pictures, marking out individual particles in three-dimensional scans of a battery electrode.

Models also run the experiment. In a closed loop, software chooses the next sample to make or measure, interprets the measurement, and updates its picture of the material system before choosing again. Related models propose how to make a target compound at all: ranking published recipes by similarity to the target, suggesting starting chemicals, and estimating a firing temperature. Training data comes from calculated databases, simulated results and figures scraped from the literature, which is why the same model is often fitted to numbers no one measured directly.

Why researchers use it

The field's bottleneck is that there are far more candidate materials than anyone can make or compute. Quantum-mechanical calculation is accurate but expensive, so it cannot be applied to every composition, every arrangement, or every system large enough to resemble a real sample. A trained model substitutes for that cost: it approximates the expensive calculation, or it approximates the expert judgement of which candidate is worth trying next.

It also stands in for human attention. Labelling thousands of particles in tomographic slices, segmenting diffraction patterns into regions of distinct crystal phase, or combing the literature for comparable syntheses are all tasks a person can do and few can do at scale. Encoding physical rules into these models, such as constraints on how many phases can coexist, is a way of keeping statistical pattern-finding tethered to known chemistry rather than letting it drift.

In the record

Limits

These models interpolate within the data they were fitted to. A potential trained on one family of structures can give confident, wrong forces outside it, and a property model fitted to simulated uptakes inherits the simulation's errors as well as its own. Literature-derived labels carry publication bias: failed syntheses are rarely reported. Image segmentation depends on manual annotations that are themselves uncertain at particle boundaries. And a prediction of what a compound should do is not evidence that it can be made, which is why experimental confirmation remains the test.

Written by AIxSci: the background from general knowledge, the examples from checked records.

What the AI was for

• Property prediction
• Simulation surrogate
• Experimental design
• Classification
• Candidate generation
• Structure determination
• Segmentation
• Detection
• Literature synthesis

Model families

• Multilayer perceptrons
• Linear models
• Random forests
• Graph neural networks
• Gradient-boosted trees
• Gaussian process
• Support vector machines
• Clustering
• Transformers
• Convolutional neural networks
• Autoencoders
• Probabilistic graphical models
• Recurrent neural networks
• Large language models
• Diffusion models
• Generative adversarial networks

Named models

• XGBoost
• MACE
• M3GNet
• LightGBM
• CHGNet
• CatBoost
• ChatGPT and GPT models

Every article