Chapter 12 — Foundation Models for Materials¶
Chapter 11 closed with a workflow that, by 2024 standards, was already the state of the art: a Bayesian-optimisation loop driving a sequence of DFT calculations, guided by a graph neural network trained on a few thousand structures. The bottleneck, as ever, was the surrogate. Each new chemical system demanded a new training set, a new model, a new round of hyperparameter sweeps. The pipeline worked, but it did not scale across the periodic table.
This chapter is about the architectural and conceptual shift that broke that ceiling: the arrival of foundation models in materials science. The term is borrowed deliberately from the language-model literature, where a single network — pre-trained once, at great expense, on a vast and heterogeneous corpus — turns out to be useful, with little or no further training, for an astonishing range of downstream tasks. GPT, CLIP and their descendants displaced an entire ecosystem of task-specific NLP models. The question this chapter asks is whether the same thing is happening, right now, in computational materials science.
The answer, as of 2026, is a qualified yes. Three lines of work make the case.
First, universal machine-learning interatomic potentials. A single MACE-MP-0, CHGNet or SevenNet model, trained on relaxation trajectories spanning most of the periodic table, can run molecular dynamics on chemistries it has never seen during training — and frequently match the accuracy of bespoke potentials fitted to that exact system. The "zero-shot universal MLIP" is now a routine starting point for any new simulation, in much the way that one starts a chemistry calculation by reaching for PBE rather than deriving a new functional.
Second, generative models for crystal structures. MatterGen, CDVAE, DiffCSP and related diffusion-based architectures invert the standard forward pipeline: instead of predicting a property from a structure, they generate candidate structures conditioned on a target property. The classical inverse-design problem, which dominated a generation of PhD theses, is being repackaged as a sampling problem in a learned manifold of stable crystals.
Third, and most speculatively, multimodal models that tie together structure, computed spectra, experimental measurements and the text of the scientific literature. The picture here is less mature, but the direction is clear: a single embedding space in which a query like "a transparent conductor with band gap above 3 eV, made of earth-abundant elements" returns a ranked list of candidate compositions together with predicted XRD patterns and citations to the closest prior work.
The chapter is organised as follows. Section 12.1 lays out the foundation-model paradigm in general terms — what it borrows from NLP and computer vision, why materials science has the structural properties (a small alphabet, a shared underlying physics) that make the paradigm viable, and how to think about zero-shot, few-shot and fine-tuning regimes in this context. We will look at the practical question that every reader will eventually face: given a small, domain-specific dataset, when does one fine-tune a universal model and when does one train from scratch?
Section 12.2 turns to MACE-MP-0 in detail. We will install the model, run a hundred-step molecular dynamics trajectory on a binary oxide with zero further training, and inspect its energy and force errors against single-point DFT. Then we will fine-tune it on a hundred structures of a domain-specific system — a perovskite catalyst, say — and observe the cost–accuracy curve. The section closes with a comparison table of the universal-MLIP zoo as it stood at the end of 2025: MACE-MP-0, CHGNet, M3GNet, SevenNet, Orb, and several smaller specialised descendants.
Section 12.3 covers generative models. We will sketch the architecture of MatterGen — a graph-based diffusion model with equivariance built in — and walk through a complete validation pipeline: generate candidate structures from the model, relax each with a universal MLIP, screen the relaxed structures by DFT, and finally hand-check the most promising against the synthesizability heuristics of Chapter 11. Limitations are discussed candidly. The model knows nothing about kinetic accessibility and cheerfully proposes phases that are thermodynamically reasonable but practically unmakeable.
Section 12.4 looks forward. We discuss autonomous laboratories (A-Lab at LBL, the Self-Driving Lab consortium), the closing of the simulation–experiment loop, and the open problems that the current generation of foundation models does not solve: long-range electrostatics, charge transfer, excited states, magnetic ordering in MLIPs, and the question of whether the universal models truly generalise across the periodic table or merely interpolate within the training distribution. A short reading list of papers from 2024–2026 points to the work most worth following.
By the end of the chapter you will have run, fine-tuned and critically assessed a universal MLIP, generated candidate structures with a diffusion model, and understood enough of the landscape to read the literature as it appears. The appendix that follows the chapter consolidates the mathematical, computational and bibliographic resources used throughout the book.
A final remark. Foundation models do not abolish the methods of Chapters 5 through 9. DFT remains the reference. Classical MD remains the workhorse for million-atom simulations. What changes is the entry point: most new projects, in 2026, begin with a pre-trained model and add specificity only where the data demand it. The art lies in knowing when to trust the foundation and when to bypass it.