Research
Five research programs.
Benchmarks and reports.
Benchmarks, reports, and technical results across MatterSpace, DLMS, CDE, and ACI.
MatterSpace Research
Generation research.
The Universal Generation Engine for Science and Engineering
MatterSpace research builds a goal-driven inverse generation engine that enforces domain rules during candidate creation, not after. Search budgets go to viable outputs across Lattice and Vital workflows instead of filtering failures.
Constrained Generative Dynamics
The Problem
Generative AI produces candidates from learned distributions, but in scientific domains 90%+ of generated outputs are physically invalid. Post-hoc filtering consumes compute without validity guarantees.
Our Approach
MatterSpace enforces physical constraints during generation, not after. Bond lengths, coordination numbers, symmetry groups, and charge neutrality are structural properties of the generation process itself.
Multi-Objective Inverse Design
The Problem
Real design problems optimize across competing objectives simultaneously. Single-objective optimization misses the diversity of solutions that engineers and scientists need to evaluate trade-offs.
Our Approach
MatterSpace uses Pareto-optimal search algorithms that maintain structured archives of candidates across competing objectives, delivering the full trade-off landscape instead of a single collapsed solution.
MatterSpace Reports
DLMS Research
Dynamic system research.
Dynamic System Learning and Modeling
DLMS research targets systems where behavior over time matters: temporal dynamics, feedback loops, and evolving system states between static discovery and deployed adaptation.
Temporal Dynamics Modeling
The Problem
Most ML models treat data as static snapshots. Real systems evolve: state at any moment depends on history, feedback, and accumulated interactions that static models cannot capture.
Our Approach
DLMS builds models that learn from temporal sequences, capturing how system state evolves and how current behavior depends on history and feedback structure.
Feedback and Stability Analysis
The Problem
Complex systems with feedback loops exhibit oscillations, instabilities, and mode transitions that simple predictive models miss entirely.
Our Approach
DLMS incorporates control-theoretic principles to model feedback, analyze stability, and predict system behavior under changing conditions. Dynamics are first-class structure, not noise.
CDE Research
Discovery research.
The Causal Dynamics Engine for Science and Engineering
The CDE research program focuses on discovering interpretable scientific structure from observational data.
Causal Structure Learning
The Problem
Correlation-based models cannot distinguish cause from effect. When gene A and protein B co-vary, standard models cannot determine directionality or confounders.
Our Approach
CDE discovers directed causal graphs from observational data. The Causal Dynamics Engine separates genuine causal edges from spurious correlations and designs targeted experiments to resolve ambiguity. CDE is patent pending.
Governing Equation Discovery
The Problem
Scientific progress depends on interpretable mathematical laws — not black-box predictors. A model that predicts a pendulum's position cannot reveal the governing equation.
Our Approach
CDE searches for closed-form mathematical laws directly from data. The output is interpretable mathematics that domain experts can read, verify, and compose.
Physics-Informed Discovery
The Problem
Treating physical data as generic inputs ignores conservation laws and symmetries, requiring more data and producing inconsistent predictions.
Our Approach
CDE embeds domain-specific physical constraints directly into the discovery process, producing physically consistent outputs by construction with less data.
Multi-Objective Search
The Problem
Scientific design problems are multi-objective and non-convex. Standard optimization collapses to a single optimum and ignores the diversity of solutions scientists need.
Our Approach
CDE maintains diverse sets of high-quality solutions across competing objectives instead of collapsing to a single optimum.
CDE Reports
ACI Research
Stability, plasticity, editability, and deployment research.
Continual Learning After Deployment
ACI research focuses on post-launch model change with three pillars: plasticity to learn efficiently without full retraining, stability against catastrophic forgetting on protected knowledge, and editability for exact unlearning.
Stability On Protected Audit Sets
The Problem
A deployed system has to change without drifting on protected outputs such as safety examples, contractual golden sets, and operational invariants.
Our Approach
ACI treats protected outputs as first-class research objects. Retained constraints can hold exactly or within a certificate-bounded guarantee depending on the update path.
Plasticity Under Bounded Resources
The Problem
Most deployments can accept new items only by waiting for a retraining cycle or by routing around the model with retrieval and middleware.
Our Approach
ACI studies how to bind new admissible items in bounded time and bounded memory so tenant facts, device habits, and policy items can enter live systems without becoming full retraining projects.
Editability And Same-Timeline Restoration
The Problem
Deletion, rollback, and policy reversal are weak points in most deployment stacks because removing a specific learned contribution is usually approximate.
Our Approach
ACI treats exact removal as a core contract. The system removes learned contributions and restores the correct state for protected scopes.
Edge And Language Attachment
The Problem
Robotics and language systems require practical deployment methods rather than one-size-fits-all claims.
Our Approach
The current research position is honest: bounded refinement and safety layers for robotics, structured-task attachment first for language systems, and higher-capacity methods only where evaluation targets exist.
ACI Benchmarks
Public Evidence
Measurable results across programs.
Key evidence across research programs — blind rediscovery, benchmark results, and foundational theory.
Research discussions.
We work with research institutions, national laboratories, and industry R&D teams. If your problem requires understanding rather than only prediction, we can discuss the appropriate technical path.