# ThinkMachine — Full Context

> This is the full inline expansion of all key content from ThinkMachine's website, provided as a single document for LLM ingestion. For the concise version, see /llms.txt.

---

## About ThinkMachine

ThinkMachine builds Equipment Agents (EA) — autonomous cyber-physical cognitive nodes that embed first-principles physics (thermodynamics, fluid mechanics, mass-transfer kinetics) into edge-compute for closed-loop industrial control.

Each EA node executes a **Perceive → Reason → Plan → Act** cognitive loop at millisecond latency, delivering physics-guaranteed perception, real-time reasoning, and optimal control without human in the loop.

**Key Metrics**: R² improvement +700%, NRMSE reduction -52% on independent industrial test sets vs. classic deep learning approaches.

**Target Industries**: Metals & smelting, chemicals & petrochemicals, biopharma & fermentation, mining & mineral processing.

---

## Agent Cognitive Stack — 5-Layer Architecture

### L0: First Principles
PDEs, ODEs, thermodynamics, fluid dynamics, reaction kinetics. The mathematical foundation that governs all industrial processes. Navier-Stokes equations, energy balance, mass conservation, population balance models.

### L1: Physics-Informed Machine Learning
PINNs (Physics-Informed Neural Networks), neural operators, physics-constrained loss functions. Embeds L0 governing equations as hard/soft constraints into neural network training, ensuring predictions satisfy conservation laws.

### L2: Equipment Agent Node (EA) — The Cognitive Core
The autonomous edge-compute node that executes the closed-loop cognitive cycle:

- **Perceive**: Multi-modal sensor fusion + PINNs-based soft sensing. Estimates unmeasurable internal states (concentration fields, temperature distributions, particle size distributions) from limited sensor inputs.
- **Reason**: Conservation-law-embedded inference engine. Performs inverse identification of key physical parameters (heat transfer coefficients, reaction rate constants, wear rates) for explainable diagnostics.
- **Plan**: Model Predictive Control (MPC) and Safe Reinforcement Learning within Control Barrier Function (CBF) safety envelope. Computes optimal control trajectories with mathematically guaranteed zero catastrophic risk.
- **Act**: Direct output to PLC/DCS actuators — no human in the loop for routine operations. Peer-to-peer multi-agent coordination protocol for cross-equipment optimization.

### L3: Multi-Agent Swarm (MAS)
Cross-equipment P2P coordination for plant-wide closed-loop optimization. Multiple EA nodes communicate via lightweight message passing to achieve global objectives (e.g., grinding-classification-flotation circuit optimization) that no single node can achieve alone.

### L4: Industrial Applications
Four primary deployment scenarios:
1. **Process Optimization**: Agent autonomously searches golden batch parameters within CBF safety boundaries via closed-loop optimization.
2. **Energy Reduction**: Agent autonomously learns energy-yield Pareto frontiers, outputting optimal operating points in real time.
3. **Predictive Maintenance**: Agent tracks physical parameter degradation online, autonomously predicts Remaining Useful Life (RUL) and triggers maintenance work orders.
4. **Operator Training**: Operators train under agent guidance with real-time physics feedback and decision recommendations.

---

## Agent Foundry — Pre-trained Equipment Agents

### Fermenter Agent
- **Physics Domain**: Cell kinetics, O₂ mass transfer, metabolic pathway modeling
- **Capability**: Autonomous metabolic pathway optimization. Real-time estimation of product concentration and respiratory quotient via cell kinetics and O₂ transfer dynamics.
- **Industry**: Biopharma, chemicals, food & beverage

### Furnace Agent
- **Physics Domain**: Multi-phase thermodynamics, radiative heat transfer, slag chemistry
- **Capability**: Real-time electrode and lining wear sensing. Safety-constrained power curve optimization via multi-phase heat/mass transfer models.
- **Industry**: Metals, advanced materials, smelting

### Hydrocyclone Agent
- **Physics Domain**: Turbulent multiphase fluid dynamics, CFD-based particle separation
- **Capability**: High-frequency feed disturbance sensing, online turbulent flow inference, dynamic underflow/overflow tuning. Zero oversize particle bypass guarantee.
- **Industry**: Mining classification, oil-water separation, mineral processing
- **Live Demo**: https://hydrocyclone-simulator.vercel.app/

---

## Why Equipment Agents (Not LLMs or Passive Simulation)

### Problem 1: LLMs Don't Apply to Industrial Control
Large Language Models suffer from hallucination, cannot enforce conservation laws (mass/energy/momentum balance), and lack real-time deterministic inference capability required for safety-critical control loops.

### Problem 2: Passive Simulation Has Peaked
Traditional CAE/CFD/FEM simulations are offline, open-loop, and require hours to days per run. They cannot close the loop with real-time control. They observe but cannot act.

### Solution: Equipment Agents
Equipment Agents embed first-principles physics directly into edge-compute neural networks, creating autonomous cognitive nodes that:
- Run at 1 kHz inference frequency (millisecond latency)
- Enforce physical conservation laws by construction
- Execute closed-loop control with CBF safety guarantees
- Require only dozens of on-site samples for calibration via transfer learning

---

## Performance Validation

| Metric | Classic Deep Learning | Equipment Agent | Improvement |
|--------|----------------------|-----------------|-------------|
| R² (coefficient of determination) | Baseline | +700% | 7× better fit |
| NRMSE (normalized error) | Baseline | -52% | Half the error |
| Inference latency | N/A | < 10 ms | Real-time capable |
| Calibration samples needed | Thousands | Dozens | Few-shot transfer |

---

## Company Information

- **Name**: ThinkMachine
- **Founded**: 2023
- **Website (CN)**: https://www.thinkmachine.work/
- **Website (EN)**: https://www.thinkmachine.work/index_en.html
- **Contact**: https://www.thinkmachine.work/#contact
- **Positioning**: Autonomous Cognition for Every Industrial Asset

---

## Contact Information

To schedule a demo or partnership inquiry:
- **Form URL**: https://www.thinkmachine.work/#contact
- **Required fields**: Full name, Company name, Business email
- **Optional field**: Description of industrial challenge / pain points