The AI-for-materials landscape in 2026: who is building what

Materials discovery went from an academic side-project to a funded race over the last two years. This post is a plain roundup of who has shipped what — no Accretion numbers, just the public record, for anyone trying to understand the field before picking a vendor or a research direction.

The model layer

None of this runs on private data alone. The Materials Project — a DOE-funded open database out of Lawrence Berkeley National Lab, computed on DOE supercomputers since 2011 — is the public foundation most of the field trains or benchmarks against. NIST built on the same idea early: its ALIGNN model (2021) trained a graph neural network on NIST's own JARVIS-DFT database plus the Materials Project, predicting over 100 material properties years before graph networks became the field's default. Google DeepMind's GNoME (Graph Networks for Materials Exploration) used the same core approach to screen candidate crystal structures at a scale no manual search could match, expanding the catalog of known stable inorganic materials by roughly an order of magnitude — from the tens of thousands into the hundreds of thousands. Microsoft Research followed with MatterGen, a generative diffusion model that designs a structure directly from a target property rather than screening a candidate list, paired with MatterSim, a simulator for evaluating those designs under realistic conditions.

The specialists

Schrödinger, longer-established in computational chemistry, has extended its physics-based simulation platform with machine learning for both materials and drug discovery — a reminder that the ML layer works best paired with a physics engine that already understands the domain. A newer cohort — Orbital Materials and CuspAI among them — is applying the same generative-plus-simulation pattern to narrower, high-stakes targets: clean energy materials and carbon capture sorbents specifically, rather than materials discovery in general.

The pattern across all of it is consistent: generative models propose, graph neural networks estimate properties cheaply, and physics-based simulation (usually DFT) checks the results that matter before anyone trusts them enough to synthesize. That three-stage loop is not any one company's secret — it is becoming the field's default architecture, which is exactly why the differentiator is shifting from "do you have a model" to "how rigorously do you validate what it proposes."

We built Accretion Core around the same loop for that reason. Where we think we differ is covered on the Platform page — this post is deliberately about everyone else.