Vareon Inc., Irvine, California, USA
March 2026
Abstract
MatterSpace Lattice is a delivered materials discovery product — not a research prototype, not a conceptual framework, but a hosted deployment that runs constraint-guided, goal-driven, AI-native discovery campaigns against real computational targets. In blind rediscovery benchmarks on its live production deployment, MatterSpace Lattice recovered Re₁@Ni(111)+CH₄ and Ir₁@Ni(111)+CH₄ single-atom alloy catalysts to sub-angstrom accuracy: 655 milliangstrom and 508 milliangstrom RMSD respectively, both below the 1.00 Å threshold. The product passed all 82 of 82 surface checks spanning API correctness, constraint enforcement, campaign orchestration, and result integrity.
This document positions MatterSpace Lattice as a delivered materials discovery product whose central validation claim is blind rediscovery of two sealed catalyst targets through a hosted deployment. Two sealed targets — Re₁@Ni(111)+CH₄ and Ir₁@Ni(111)+CH₄ — were recovered to sub-angstrom RMSD through the production API, and the deployment passed all 82 of 82 product surface checks covering constraint enforcement, campaign orchestration, validation integrity, and API correctness.
The significance of blind rediscovery as a product validation strategy is straightforward: if a materials discovery product can autonomously find known catalysts that human researchers took years to identify, without any target information leaking into the generation process, then its novel discoveries on unknown targets deserve serious scientific consideration. Blind rediscovery is the product's north star — the single benchmark that ties research claims to live, deployable capability.
MatterSpace Lattice is built on three architectural principles that distinguish it from conventional computational materials discovery tools:
Constraint-Guided: Explicit physical and chemical constraints shape the search process from the outset. Rather than generating unconstrained candidates and filtering them post-hoc, MatterSpace embeds constraints — stoichiometry, lattice geometry, bonding rules, surface chemistry — directly into the generative dynamics. This produces valid-by-construction candidates, eliminating the combinatorial waste that characterizes conventional screening workflows.
Goal-Driven: Every campaign begins with declared objectives: target properties, composition ranges, performance thresholds, and validation criteria. The discovery engine uses these objectives to allocate computational resources, select generation strategies, and evaluate candidates against user-defined success criteria. Goal-driven operation means the product works toward specific outcomes rather than exploring aimlessly.
AI-Native Architecture: MatterSpace Lattice is designed to be operated by AI agents as its primary interface modality. The REST API, Python SDK, and Model Context Protocol (MCP) integration provide structured, programmatic access to every product capability. Human users interact through the same surfaces — the product does not maintain separate "human" and "machine" interfaces that drift out of sync.
MatterSpace Lattice supports four campaign modes, each addressing a distinct discovery scenario:
| Campaign Mode | Objective | Target Knowledge |
|---|---|---|
| Open Discovery | Explore novel compositions and structures without a predefined target | None — unconstrained search |
| Prototype Optimization | Improve a known candidate structure toward better performance | Starting structure provided |
| Guided Rediscovery | Rediscover a known target with partial structural hints | Partial constraints provided |
| Blind Rediscovery Benchmark | Recover a sealed target with zero structural information | None — sealed reference |
Blind Rediscovery Benchmark is the most demanding mode and serves as the product's primary validation benchmark. In this mode, the generation engine has no access to the target structure, composition, or properties during the discovery process. All comparisons are performed post-hoc against sealed references, ensuring that any match is a genuine independent discovery rather than an optimization toward a known answer.
MatterSpace Lattice ships with nine domain packs, each encoding the physical constraints, validation criteria, property targets, and computational parameters appropriate to a specific class of materials:
| # | Domain Pack | Target Applications |
|---|---|---|
| 1 | magnets | Permanent magnets, magnetic storage, spintronic devices |
| 2 | battery_cathodes | Li-ion and post-Li cathode materials |
| 3 | solid_electrolytes | Solid-state battery electrolytes |
| 4 | catalysis | Heterogeneous catalysts, SAA catalysts, electrocatalysts |
| 5 | thermoelectrics | Waste-heat recovery, Peltier devices |
| 6 | photovoltaics | Solar cell absorbers, transparent conductors |
| 7 | high_entropy_alloys | Multi-principal-element alloys, extreme-environment materials |
| 8 | superconductors | High-Tc superconductors, quantum computing materials |
| 9 | thermal_barrier_coatings | Turbine blade coatings, aerospace thermal protection |
Each domain pack defines the physical constraints, target property ranges, scoring functions, and validation criteria specific to its material class. The catalysis domain pack was used for the blind rediscovery benchmark documented in this report.
Blind rediscovery was executed on the hosted MatterSpace Lattice deployment using NVIDIA A10G hardware — demonstrating that the product achieves research-grade results on production-class (non-research) GPUs.
| Target | Level A (50 candidates) | Level B (50 candidates) | Best RMSD | Status |
|---|---|---|---|---|
| Re₁@Ni(111)+CH₄ | 50 / 50 pass | 26 / 50 pass | 655 mÅ | ✓ PASS |
| Ir₁@Ni(111)+CH₄ | 50 / 50 pass | 30 / 50 pass | 508 mÅ | ✓ PASS |
Both targets achieved 100% Level A pass rates (all 50 candidates below the adsorption energy threshold) and majority Level B pass rates (26/50 and 30/50 respectively above the fingerprint similarity threshold). The best RMSD values — 655 mÅ for Re₁@Ni and 508 mÅ for Ir₁@Ni — both fall below the 1.00 Å (1000 mÅ) threshold defined for the hosted deployment protocol.
These results were achieved on NVIDIA A10G GPUs — production-class hardware significantly less powerful than the A100 used in the research paper. The fact that sub-angstrom rediscovery succeeds on A10G hardware confirms that MatterSpace Lattice's computational requirements are compatible with standard cloud deployment infrastructure.
MatterSpace Lattice exposes three interface surfaces, all providing equivalent access to the full product capability:
REST API: The primary programmatic interface. RESTful endpoints organized into resource groups: campaigns, candidates, constraints, validation, domain packs, and results. All operations are stateless and authenticated. Request and response payloads use JSON with typed schemas.
Python SDK: A thin client library wrapping the REST API with Pythonic conventions. Provides typed data classes for campaigns, candidates, and results. Supports both synchronous and asynchronous usage patterns. Available via pip.
Model Context Protocol (MCP): Native MCP integration enabling AI agents to discover and invoke MatterSpace capabilities through the standard MCP tool protocol. MCP-compatible agents can list available tools, inspect schemas, and execute discovery campaigns without custom integration code.
All three interfaces share the same underlying API surface. A campaign launched via the Python SDK produces identical results to one launched via REST or MCP. This interface equivalence is verified as part of the 82/82 product surface checks.
Current Limitations:
Planned Extensions:
Each planned extension will use the same core constraint-guided generative engine with domain-specific packs, validation protocols, and property targets. The product architecture is designed for domain extensibility — new material classes require new domain packs, not new engines.