AI for material discovery: reinforcement learning by Park and Walsh

A paper published in Nature Machine Intelligence by Park and Walsh describes the application of AI and reinforcement learning to discover new crystalline materials. The authors applied this approach—a framework for targeted search of thermodynamically stable structures with desired properties. This is one of the new ways to accelerate scientific discovery through machine learning.

How Reinforcement Learning Searches for Crystals

Crystalline materials are built on regular three-dimensional structures of atoms, where each atom occupies a specific location in the lattice. The space of possible structures is enormous: theoretically, there are billions of combinations of atom arrangements of different elements.

Checking every option computationally is expensive—for each potential crystal, you need to calculate the energy of the structure through quantum simulations, verify thermodynamic stability, and assess functional properties. Reinforcement learning fundamentally changes the strategy: instead of random search, the algorithm learns from results and receives "rewards" for discovered materials that are simultaneously low-energy and have the desired properties—like a game AI that improves through playing thousands of matches.

Why This Matters for Technology

New crystalline materials determine the development of high-tech industry:

• Batteries and supercapacitors for electric vehicles
• Solar panels with high efficiency
• Semiconductors for microelectronics and quantum computers
• Catalysts for fertilizer synthesis, hydrogen production, and CO₂ capture

Each breakthrough in materials science requires years of experiments. Accelerating materials search means accelerating technological progress—Park and Walsh show that RL can narrow the search space, directing it from random variants to promising combinations never before considered.

AI for Scientific Discovery

Publication in Nature Machine Intelligence—one of the top journals at the intersection of ML and science. This is part of a recognized trend: transition from "AI for data analysis" to "AI for discovery." Before Park and Walsh, deep learning was already applied to molecular design (DeepMind's AlphaFold for protein structure, generative models for organic molecules). The authors extend this practice to the space of crystalline structures, where combinatorics is even more complex.

Particularly significant: Park and Walsh orient RL not just toward finding one optimal material, but toward discovering a diverse set of variants—which is more practical for experimenters receiving a corridor of alternatives.

What This Means

Material discovery through RL becomes part of the standard AI toolkit for science alongside language models for hypothesis generation and simulations for validation. Immediate applications: accelerating R&D in batteries, semiconductors, solar panels. Distant horizon—automated laboratories where RL-driven search proceeds in a cycle with robotic experiments: algorithm proposes → robot validates → data returns to the model.

How

Does Reinforcement Learning Work for Discovering New Materials? It is a framework for targeted search of thermodynamically stable structures, where each atom of the crystal occupies a specific location in a regular three-dimensional lattice.