Every material that changed the world — lithium-ion cathodes, carbon fiber, high-temperature superconductors — took years of trial and error to find. Not because the answer wasn't there. Because experiments are slow, expensive, and sequential. One researcher. One hypothesis. One result. Repeat.
That is the bottleneck. And it is not fundamental.
The research process is the constraint
Material science has never been limited by ambition. Researchers have always known what they wanted: stronger alloys, denser batteries, lighter polymers. The problem is the loop. Forming a hypothesis takes weeks of literature review. Running a simulation requires access to supercomputing clusters — and weeks more of compute time. Interpreting results is manual. Iterating is slow. Most paths lead nowhere, and the feedback that could redirect you arrives months later.
The average time from discovery to validated material is measured in years. Sometimes decades.
This is an engineering problem, not a scientific one. The rate of discovery is bounded by throughput, not by the quality of the underlying physics.
What we're building
Supermatter builds SM-1 — an AI material scientist that runs material science and material engineering workloads autonomously, end-to-end.
SM-1 does not assist researchers. It conducts research. It reads the literature, forms hypotheses, designs experiments, runs them inside quantum-accurate virtual laboratories, measures the results, eliminates dead ends, and recurses. Around the clock, without pause.
The inputs are engineering goals: design a cathode material with higher lithium diffusivity and lower volume expansion than NMC811, or find a structural polymer that maintains tensile strength above 200 MPa at 180°C. SM-1 does not search for existing materials that partially match. It designs new ones — proposing novel crystal structures, evaluating their properties from first principles, and iterating toward the specification.
The outputs are validated candidates, quantum-to-multiscale simulation pathways, and a full chain of reasoning — every hypothesis explored, every measurement taken, every decision traceable.
AI material discovery
The dominant paradigm in computational materials science for the past decade has been screening: take a large database of known materials, compute properties, filter. This works well when the answer is already in the database. It fails when the answer has never been synthesized before.
Most commercially valuable materials are not in any database.
SM-1 operates in generative mode. Rather than screening candidates, it proposes novel crystal structures — lattice geometries, atomic compositions, and bonding configurations that do not yet exist — and evaluates them through quantum-accurate simulation. Structure prediction is guided by the research goal, constrained by synthesizability, and refined through iterative feedback from virtual experiments.
This shifts the question from which known material is closest? to what is the best possible material for this application, and how do we make it?
Battery and energy materials
Battery development is one of the highest-leverage applications of AI-driven materials research. The performance ceiling of lithium-ion technology is largely determined by three materials: the cathode, the anode, and the solid electrolyte. Each has a multi-dimensional optimization landscape — energy density, cycle life, thermal stability, ionic conductivity, cost — and the tradeoffs between them are deeply non-linear.
SM-1 is designed for exactly this class of problem. It can explore cathode chemistries beyond the NMC and LFP families that dominate today, searching for layered oxides, polyanionic compounds, and disordered rocksalt structures that offer better tradeoffs under specific operating conditions. It can evaluate solid electrolyte candidates for ionic conductivity and electrochemical stability simultaneously. It can model anode expansion behavior over hundreds of simulated charge cycles.
The goal is not incremental improvement. It is identifying the next generation of energy materials before the competition does.
Structural and mechanical engineering
Material engineering problems are often framed as constraint satisfaction: achieve a target strength, stiffness, or thermal conductivity while minimizing weight, cost, or processing complexity. These constraints interact in ways that are difficult to reason about analytically — which is why physical prototyping has historically been unavoidable.
SM-1 closes the loop computationally. For structural alloys, it can predict yield strength, fatigue resistance, and corrosion behavior as a function of composition and microstructure. For polymer composites, it can model viscoelastic behavior, thermal degradation, and fiber-matrix adhesion. For ceramics, it can predict fracture toughness and thermal shock resistance without a single physical test.
The multiscale bridge SM-1 produces is not a theoretical output. It spans from quantum-accurate electronic structure to bulk engineering performance — temperature profiles, alloying sequences, and processing conditions grounded in the simulated thermodynamics and kinetics of the system.
Computational methods: DFT, GNN, and molecular dynamics
The simulation stack inside SM-1's virtual laboratories combines three methodologies, each operating at its appropriate scale.
Density functional theory (DFT) handles electronic structure: formation energies, band gaps, magnetic moments, and charge density distributions. These are the ground-truth properties that determine whether a candidate material is thermodynamically stable and worth exploring further. DFT calculations are computationally expensive, so SM-1 uses them selectively — at decision points where high-confidence property estimates matter.
Graph neural networks (GNNs) trained on DFT datasets provide fast approximate property predictions across the compositional space. Where a DFT calculation might take hours, a GNN inference takes milliseconds. SM-1 uses GNN predictions to navigate the hypothesis tree efficiently, reserving DFT validation for the most promising candidates.
Molecular dynamics (MD) simulates material behavior over time — thermal diffusion, grain boundary migration, defect formation and propagation, ion transport. This is where properties like ionic conductivity, creep resistance, and cycle degradation are evaluated. MD connects the static structure predictions from DFT and GNNs to the dynamic performance metrics that engineering applications actually require.
The combination of all three — GNN for rapid exploration, DFT for ground-truth validation, MD for dynamic performance — gives SM-1 the coverage to evaluate a candidate material from atomic stability to real-world operating conditions within a single investigation.
Multi-scale fluency
One of the hardest problems in materials research is working across scales. Atomic structure determines microstructure. Microstructure determines bulk properties. These scales are deeply coupled, but human researchers almost always specialize in one.
SM-1 operates across all three simultaneously. It models electronic structure at the quantum level, predicts grain boundary formation and defect propagation at the microstructural level, and maps those predictions to macro-scale mechanical and thermal performance. A single investigation can span from electron orbital geometry to tensile yield strength without losing resolution at any layer.
A frontier agentic system
SM-1 is not a tool you prompt. It is an autonomous agent that runs research sessions — continuously, for hours or days at a time — without human intervention between steps.
Most AI systems operate in request-response cycles. You send a query, you get an answer, you send the next query. The reasoning is stateless. The context resets. This is fine for question answering. It is inadequate for research, where the value compounds across dozens of interconnected experiments and the most important decisions are made ten steps into an investigation.
SM-1 maintains persistent research sessions. A single session can span hundreds of simulation runs, thousands of property evaluations, and multiple recursive hypothesis revisions — all running natively inside your environment, against your data, with full access to your existing experimental results, proprietary databases, and internal literature. The system reads your context at the start and carries it through the entire investigation.
Sessions are not black boxes. You can inspect the current hypothesis tree at any point, redirect the investigation, inject constraints, or terminate early and take the partial results. The agent's working memory — every observation, every intermediate conclusion — is structured and queryable throughout.
Bring your own compute
SM-1 is designed to run where your data lives.
For teams with existing HPC infrastructure, SM-1 deploys directly onto your compute cluster. The simulation workloads — DFT jobs, molecular dynamics runs, GNN inference — execute on your hardware, inside your security perimeter, without data leaving your environment. This matters for organizations working on proprietary material formulations, confidential R&D programs, or regulated research contexts.
The simulation stack is built on industry-standard frameworks. NVIDIA CUDA for GPU-accelerated quantum chemistry and molecular dynamics. Triton for custom kernel optimization on high-throughput inference workloads. PyTorch and JAX for the GNN property prediction models — JAX in particular for the differentiable physics components, where gradients through the simulation are used to guide structure optimization directly.
This means SM-1 integrates into existing MLOps pipelines rather than replacing them. Teams running internal DFT workflows, custom force fields, or proprietary simulation codes can connect those systems as callable tools that SM-1 orchestrates alongside its built-in methods. The agent adapts to your stack.
Computational materials science has long suffered from a credibility problem: predicted materials that cannot actually be made. A material can be thermodynamically stable on paper and completely unsynthesizable in practice — because the synthesis route doesn't exist, the required precursors are unavailable, or the processing conditions are beyond what any lab can achieve.
SM-1 treats synthesizability as a hard constraint, not an afterthought. Candidate structures are evaluated not only for their target properties but for their stability under realistic synthesis conditions — temperature, pressure, atmosphere, precursor chemistry. The quantum-to-multiscale bridge SM-1 produces is not a suggestion. It is a grounded simulation pathway, derived from the same stack that validated the material's properties at every scale.
This is the difference between a theoretical prediction and a practical discovery.
No lab required
The most immediate consequence of virtual research is that physical manufacturing becomes optional at the discovery stage.
Prediction without manufacturing. Material properties — yield strength, ionic conductivity, band gap, thermal stability — are predicted from first principles before a single gram is synthesized. Physical testing becomes confirmation, not discovery. The expensive step moves to the end of the pipeline, after the search space has already been narrowed to high-confidence candidates.
Iterative design without a lab. A design iteration that would take months of physical prototyping — synthesize, characterize, adjust composition, repeat — happens in hours inside SM-1's virtual laboratories. Failure is free. The cost of exploring a wrong direction is compute time, not materials, equipment, and researcher effort. This changes the economics of early-stage R&D fundamentally.
Parallel hypothesis exploration. Where a human research team explores one or two candidate materials at a time, SM-1 evaluates hundreds in a single session — branching across compositional spaces, pruning dead ends, and converging on the strongest directions simultaneously. The throughput advantage compounds: more candidates evaluated means a higher probability of finding the optimal material, not just an adequate one.
Cost-efficient optimization. Compute scales with problem complexity. Routine property evaluations are cheap. Hard problems — complex crystal structures, multi-constraint optimization, rare failure mode prediction — get more resources automatically. The result is orders-of-magnitude lower cost per validated candidate compared to traditional physical R&D pipelines.
Interpretability is not optional
In science, a result you cannot explain is not a result. It is a liability.
Every decision SM-1 makes is visible. Which hypotheses it explored. Why it eliminated certain candidates. What the simulation measured. How it arrived at a conclusion. The full chain of reasoning is available — not as a log file, but as structured scientific output you can interrogate, verify, and build on.
We believe interpretability is not a feature. It is the foundation that makes autonomous research scientifically credible.
Why now
Three things converged.
Quantum-accurate graph neural networks can now predict material properties — formation energy, band gap, elastic moduli, ionic conductivity — with fidelity that closes the gap between simulation and physical experiment. Virtual laboratories have reached a level of physical realism where experiments run inside them produce meaningful results. And large-scale foundation models can reason coherently over scientific literature, experimental data, and structured material databases simultaneously.
None of these existed at sufficient quality even three years ago. The integration — a closed loop from hypothesis to virtual experiment to measurement to next hypothesis — is now possible. That loop is SM-1.
What's next
We are in early access. If you are working in battery technology, polymer science, structural alloys, semiconductor design, or catalyst development — and you want to run your first autonomous research session — reach out.
We will publish our research here as we build: architectures, benchmark results, experimental comparisons, and the hard problems we run into. Starting now.