600 candidates. 23 dopant elements. Zero target knowledge. Both Re₁@Ni and Ir₁@Ni catalysts blindly rediscovered to sub-angstrom accuracy. Every candidate valid by construction.
0.408 Å
Full RMSD
Level C PASS
97.5–99%
Structural validity
By construction
~$15
Cloud cost
Single A100, 4.7 hrs
Vareon Inc. · Vareon Limited · March 2026
Materials discovery today follows a propose-then-filter paradigm. Generate millions of candidate structures randomly or through exhaustive enumeration. Evaluate them with expensive density functional theory (DFT) or machine-learned interatomic potentials. Discard the vast majority because they are physically invalid or chemically unreasonable.
The waste is staggering. Most compute cycles are spent evaluating structures that should never have been proposed. GNoME identified 2.2 million stable crystals through brute-force screening. MatterGen generates crystals with diffusion but filters for validity afterward. Open Catalyst built massive datasets for screening. CDVAE, DiffCSP, FlowMM — all rely on post-hoc filtering for physical validity.
None of these approaches can guarantee that a generated structure satisfies physical, chemical, and geometric constraints during generation. None have demonstrated blind rediscovery — starting from zero knowledge and independently finding known materials to sub-angstrom accuracy.
Inspired by control systems theory applied to generative AI, MatterSpace embeds physical, chemical, and geometric constraints directly into the generative dynamics loop. Goals and constraints are not applied after generation — they are baked into every single step of the generation process, resulting in near-guaranteed validity every time.
A SchNet-style message-passing neural network (~250K parameters) trained via Contrastive Force Matching. Operates in two modes — providing approximate forces to guide generative dynamics and serving as a learned energy landscape for adaptive exploration. Fast enough for real-time inference, accurate enough to navigate complex energy surfaces.
The constraint enforcement core. Control Barrier Functions from control systems theory, solved through quadratic programming at every dynamics step. The QP enforces that every structural constraint — minimum distances, coordination bounds, height limits — is satisfied at every single step. Solve time: under 1 ms.
Deterministic descent, Langevin stochastic exploration, quantum-inspired tunneling, and quench refinement — selected adaptively based on real-time landscape metrics. The system autonomously balances exploration and exploitation without manual scheduling.
CHGNet, MACE, or any future MLIP plugs in as a post-generation refinement calculator. The two-pass refinement protocol (coarse + fine) improved accuracy from 4.05 Å to 0.408 Å without modifying the core generative architecture.
The paradigm shift
Current generative AI suffers from the generate-and-filter approach: produce candidates blindly, then discard the invalid ones. MatterSpace's physics-inspired architecture, coupled with control theory inside the generation loop, eliminates this waste entirely. The engine generates with goals and constraints baked in, resulting in near-guaranteed validity every time. This is not a marginal improvement — it is a fundamentally different paradigm for generative AI.
MatterSpace was tested through a blind rediscovery experiment: starting from a palette of 23 dopant elements with zero target information, the engine had to independently generate candidates that match known Re₁@Ni and Ir₁@Ni single-atom alloy catalysts for methane cracking. A three-level post-hoc validation protocol measured performance, structural motif similarity, and exact geometric accuracy.
Primary Metric
581 / 600 candidates below -1.3 eV
Key Finding
Best adsorption energy: -34.73 eV
The vast majority of generated candidates exhibit strongly favorable surface binding, confirming that MatterSpace generates catalytically relevant materials — not random structures.
Primary Metric
75 matches, best similarity 0.814
Key Finding
Both Re and Ir independently identified from 23 elements
Without any target information, the engine correctly identified both target dopant elements from a 23-element palette. The probability of randomly selecting both correct elements is 0.19%. MatterSpace achieves this through learned d-orbital hybridization and favorable binding preferences.
Primary Metric
0.408 Å full RMSD (metal-only: 0.363 Å)
Key Finding
Both targets independently below 0.5 Å threshold
Ir₁@Ni at 0.408 Å and Re₁@Ni at 0.466 Å — both independently rediscovered to sub-angstrom precision. This is the first demonstration of blind generative material rediscovery achieving all three validation levels for surface catalysts.
30,000
QP solves
0.04%
Infeasibility rate
0.7 ms
Avg QP solve time
3.8%
Constraint overhead
MatterSpace is the only system achieving all three validation levels. Existing generative models demonstrate Level A capability (favorable properties) but have not demonstrated Level B (motif matching) or Level C (sub-angstrom structural reproduction) — because they are not designed for blind rediscovery.
| System | Method | A | B | C | Constraints |
|---|---|---|---|---|---|
| GNoME | MLFF screening | PASS | — | — | Post-hoc |
| MatterGen | Diffusion model | PASS | — | — | Post-hoc |
| Open Catalyst | Large-scale MLFF | PASS | — | — | Post-hoc |
| CDVAE | VAE + diffusion | PASS | — | — | Post-hoc |
| DiffCSP | Diffusion crystal | PASS | — | — | Post-hoc |
| FlowMM | Riemannian flow | PASS | — | — | Post-hoc |
| MatterSpace | CBF/QP + CHGNet | PASS | PASS | PASS | By construction |
This result was produced by MatterSpace Lattice — the materials discovery engine. But the core architecture that made it possible is domain-agnostic. The constraint enforcement, the adaptive dynamics, the quality-diversity search — none of it is specific to materials. Only the domain pack changes.
Control Barrier Functions enforcing constraints at every dynamics step (valid by construction)
Adaptive Dynamic Control for landscape navigation
Quality-diversity search maintaining diverse Pareto-optimal archives
Modular architecture — plug in any model as a refinement calculator
Multi-objective optimization toward user-defined goals
Scientists direct their agents with what they are after. Their agents pick parameters from a parameter pool curated for each pack, set constraints and goals, and MatterSpace begins generating. The models are based on strong open-source foundations — and customers can bring their own models and data too.
But that is rarely the bottleneck. Most open-source models are already strong. What's lacking is strong engineering to steer generation toward the desired solution space with small compute budgets and achieve near 100% validity by construction.
Batteries, catalysts, superconductors, magnets, photovoltaics, thermoelectrics, HEAs, electrolytes, coatings
ADMET constraints, binding affinity objectives, molecular stability. The same constraint enforcement produces valid drug candidates.
Complexity bounds, correctness constraints, optimality objectives. Valid-by-construction algorithms.
Design rule constraints, power-performance-area objectives. Physical layout validity by construction.
Expected impact
The 97.5–99% structural validity demonstrated on materials is expected to bring a new paradigm to generative AI across drug discovery, longevity research, advanced materials, and algorithm design. The physics-inspired architecture of MatterSpace — with control theory inside the generation loop — eliminates the generate-and-filter waste that plagues current generative approaches. When constraints are baked into generation, nearly every GPU cycle produces a viable candidate.
The complete peer-reviewed manuscript with all methods, results, constraint enforcement statistics, and comparison tables.
Read the Paper