MIT researchers just lifted crystal stability from under 1 percent to 68 percent by teaching AI the periodic table’s rules instead of screening junk outputs.
This is not an incremental improvement. It is a regime change. The brute-force screening era in computational materials discovery is over. On July 26, 2025, a team led by Heather J. Kulik and Mingda Li at MIT posted a preprint that achieved what the field has chased for years: a generative model that produces stable crystal structures directly, without needing to filter out the 99 percent of candidates that would fail in the real world.
The number that rewrites the economics
CrysVCD—short for Crystal generator with Valence-Constrained Design—hits 68 percent phonon stability and 85 percent thermodynamic stability when fine-tuned on stability metrics. The baseline for generative models without post-generation screening is below 1 percent.
Those numbers compress a 99 percent failure rate into a 68 percent success rate. Every materials R&D lab with a compute budget now faces a math problem, not a judgment call.
The paper is in press at Nature Computational Science as of May 2026. The code is open-source under an MIT License, sitting on GitHub. That fork count will not stay at one for long.
Why screening was a dead end
Generative models for crystal discovery have been data-hungry machines that produce mostly garbage. A diffusion model or a variational autoencoder would spit out millions of candidate structures. Then a separate screening step—density functional theory calculations, phonon dispersion checks—would discard the vast majority. The compute spent on generating junk was wasted. The compute spent on validating junk was wasted. The only thing that mattered was the tiny fraction that survived.
This pipeline made a kind of grim economic sense when there was no alternative. You paid the screening tax because you had to. But the tax was enormous, and it meant that pure data-driven generative models were computationally competitive only for the largest labs with the deepest pockets.
The old way: generate, then filter. The new way: build the rules into the generator and skip the filter.
The plugin, not the pipeline
CrysVCD works in two stages. First, a transformer-based elemental language model generates compositions that are valence-balanced—meaning the oxidation states of the constituent elements sum to zero. This is basic chemistry, the kind of rule a first-year undergraduate learns. But previous generative models did not enforce it during generation. They learned it imperfectly from data, if at all.
Second, a diffusion model takes those valence-balanced compositions and generates the crystal structures. The backbone is built on DiffCSP. Evaluation runs on MatterSim.
The key insight is that the valence constraint is embedded in the generative process itself. It is not a filter applied after the fact. It is a physical law wired into the architecture.
Mingda Li put it plainly to MIT News: "You can plug this into any kind of model, not only existing diffusion models but also future models, where people can't generate enough stable materials, and it can improve stability."
That word "plug" matters. CrysVCD is not a monolithic system that demands you rip out your existing pipeline. It is a plugin. The valence-constrained design layer can slot into diverse generative architectures. The efficiency gain is orders of magnitude in chemical valence checking compared to pure data-driven approaches with post-screening.
The end of brute-force screening
The mechanism is simple, and that is what makes it devastating to the old way. By embedding oxidation state balance directly into generation, CrysVCD eliminates the need for brute-force post-screening. The physics does the work that compute previously had to brute-force.
What this forces next is a wholesale reallocation of compute budgets across the materials discovery industry. Every major pipeline—from battery cathode startups to semiconductor fabs—must either adopt a valence-constrained generation layer or watch its compute-to-stable-material yield become economically uncompetitive. The math is unforgiving: a 99 percent failure rate versus a 68 percent success rate. That gap is wide enough to swallow a competitor whole.
Pure data-driven generative models that rely on post-hoc filtering will be relegated to exploratory tasks. They will still generate novel candidates, but no one will bet a production pipeline on them. The screening tax is now optional, and optional taxes do not get paid.
The groups that own the IP—MIT's Kulik and Li labs—have a head start. The CrysVCD code is open-source, but the expertise to extend it, fine-tune it for specific materials classes, and integrate it into proprietary pipelines sits with the people who built it. Expect at least two VC-backed materials startups to form around CrysVCD derivatives by mid-2027. The plugin architecture makes it a natural foundation for a platform play: license the integration layer, train on customer-specific property targets, collect the royalties.
Here is the predictive call, specific and falsifiable: within 12 to 24 months, adoption of valence-constrained generation will shift from a research advantage to a competitive necessity. Any lab that has not integrated physics-informed constraints into its generative pipeline by mid-2027 will be burning compute on junk while its competitors print stable candidates. If we do not see at least one major materials company publicly announce a shift to constrained-generation methods by Q3 2027, this thesis is wrong.
Reallocate your compute budget now
The operator takeaway is blunt. If you run a materials R&D lab, you need to reallocate compute from post-generation validation to pre-generation design. The CrysVCD code is open-source under an MIT License. The plugin architecture means you can integrate it into your existing pipeline without a rebuild. The cost of not adopting is a 99 percent failure rate. The cost of adopting is a GitHub clone and some integration engineering.
This is not a research paper you read and file away. It is a tool that changes the unit economics of your compute spend. The labs that move first will build up a library of valence-constrained models tuned to their specific materials classes—solid electrolytes, thermoelectrics, high-κ dielectrics—while the labs that wait will be playing catch-up with an inferior cost structure.
CrysVCD already supports conditional generation of functional materials. The paper demonstrates discovery of candidates for high thermal conductivity semiconductors and high-κ dielectric compounds. These are not academic toy problems. They are the materials that go into chips, batteries, and power electronics.
The periodic table has been taught
The brute-force era is over. The periodic table's rules have been taught to AI, and the result is a leap from near-zero to 68 percent in the fraction of generated materials that actually work. The real discovery begins now.
