UMST Prototype: Physics-Gated AI for Material Design

Overview

This repository contains the official implementation of the Differentiable Unified Material-State Tensor Optimization (DUMSTO) framework.

DUMSTO is a hybrid AI architecture that integrates rigorous thermodynamic constraints into deep learning pipelines. By enforcing the Clausius-Duhem inequality as a hard gate, DUMSTO ensures 100% thermodynamic admissibility for all material predictions and designs, bridging the gap between data-driven flexibility and physics-based safety.

Key Results (Single Source of Truth)

All results are verifiable against the Single Source of Truth (SSOT) benchmark logs located in results/ssot/. These metrics cover 16,146 real-world samples across four distribution-shifted datasets.

Method	D1 Accuracy (MAE)	Global Safety (Admissibility)	Inference Speed
DUMSTO-Hybrid (Ours)	2.99 MPa	100.0% (Guaranteed)	< 0.01 ms
XGBoost (Unconstrained)	3.05 MPa	98.0% (2% viol.)	2.7 µs
H-PINN (Hard Constr.)	8.33 MPa	88.7% (11% viol.)	5.7 µs
GNN (MatGL)	9.48 MPa	97.4% (2.6% viol.)	1.6 µs

Generative Creativity: The DUMSTO-PPO agent, operating under the constitutional gate, discovered 61 distinct Pareto-optimal designs (Yield) across 9 SCM coverage regimes, achieving a coverage score of 0.678—significantly outperforming unconstrained evolutionary baselines.

Architecture

1. Unified Material-State Tensor (UMST): A sparse, multi-scale representation ($\mathbb{R}^{64}$) encoding Material, Physics, Process, Environment, and Time states.
2. Physics Kernel (Rust): A high-performance, differentiable engine implementing 16 constitutive laws (Powers, hydration kinetics, gel-space ratio).
3. Thermodynamic Gate: A "mathematical firewall" that rejects any state transition violating entropy production constraints ($\mathcal{D}_{\text{int}} \ge 0$).
4. DUMSTO-PPO: A constitutional reinforcement learning agent with 6 reward modes (Balance, Strength, Sustainability, Durability, Cost, Printability).

Repository Structure

umst-prototype/
├── data/                   # Verified Datasets (D1-D4)
├── docs/                   # Detailed Documentation
├── results/
│   ├── ssot/               # Single Source of Truth JSONs
│   └── plots/              # Generated Figures
├── scripts/                # Python Training & Analysis Scripts
│   ├── benchmark_predictive.py
│   ├── benchmark_generative.sh
│   ├── calibrate_physics.py
│   └── ...
├── src/
│   └── rust/               # Core Physics Kernel (Rust)
│       └── core/src/tensors/kleisli.rs  # KleisliArrow admissibility monad
├── ros2_bridge/            # ROS2 bridge (Python nodes → REST/WS → Rust gate)
└── tools/                  # Verification & Setup Utilities

Quick Start

Prerequisites

• Python 3.10+
• Rust (Cargo 1.75+) for the physics kernel.

Setup

# 1. Install Python dependencies
pip install -r requirements.txt

# 2. Build Rust Kernel
cd src/rust/core
cargo build --release

Running Benchmarks

To reproduce the Predictive Power results (Table 2):

python scripts/benchmark_predictive.py

Output: results/ssot/fair_comparison_2026-01-28.json

To reproduce the Generative Design results (Table 4 & Figure 5):

./scripts/benchmark_generative.sh

Output: results/ssot/design_benchmark_latest.json

Gate Server & ROS2 Bridge

cargo run --release --bin gate_server   # REST gate on port 8765

The ros2_bridge/ directory provides a ROS2 Python package bridging the gate server to ROS2 topics. See ros2_bridge/README.md.

Datasets

We utilize a Composite Global Benchmark ($N=16,146$) merging four sources:

• D1 (UCI Concrete): The canonical compressive strength dataset ($N=1,030$).
• D2 (Zenodo NDT): Non-destructive testing data ($N=4,891$). (License: CC-BY 4.0)
• D3 (Zenodo Sun): Solar reflectance and thermal mass ($N=2,780$).
• D4 (Zenodo RH): Relative humidity / curing data ($N=7,445$).

All data is pre-processed and located in data/.

Research Publications

For the complete scientific and mathematical foundations of UMST:

• Scientific and Mathematical Foundations (71 pages)
  Complete theoretical derivation, thermodynamic proofs, and empirical validation across 16,146 samples.

• Technical Summary (8 pages)
  Concise overview of the hybrid architecture, physics-gated AI approach, and benchmark results.

Citation

If you use this code or dataset, please cite:

@software{dumsto2026,
  title={UMST: Unified Material-State Tensors for Physics-Gated AI},
  author={Shyamsundar, Santhosh and Prabhu, S. and Studio Tyto},
  year={2026},
  url={https://github.com/studiotyto/umst-prototype}
}

License

MIT License. See LICENSE for details.