Micro AI directly on Chips

onchipAI uses AI in chips, bringing micro AI capabilities to instantly react with variable values.

Design with AI establishes a more general concept that can serve multi-factory scenarios across 4 phases:

  • Phase 1: Route Selection (Milling → Mixing → Granulation → Drying → Compaction).
  • Phase 2: Teacher training via teaching how varying machine settings affect empirical qualities.
  • Phase 3: Inverse Predictor implementation (using multi-objective optimization to calculate the exact machine settings required to achieve target qualities).
  • Phase 4: Student Model Training (Knowledge Distillation) which includes loading AI in factory chips.

Our sample onchipAI solution for Pharmaceutical production

Phase 1: Route Drug: Ibuprofen Tool: RDKit Action: Parse SMILES Mill → Mix → Gran → Dry → Compaction Phase 2: Training Tool: PyTorch Geometric Model: Teacher GNN Action: 2-Stage Physics Opt. Output: Empirical Quality Predictor Phase 3: Inverse Tool: SciPy (Nelder-Mead) Goal: Hardness Target 80N Action: MQTT Telemetry Output: Optimal Force (10-25 kN) Phase 4: Edge AI Tool: XGBoost + ESP-IDF Hardware: ESP32-S3 Action: Model Distillation Output: Closed-Loop Inference < 10ms

Deployment Architecture

DeepChem Route Oracle ↓ Chosen Route: Wet Granulation Milling Mixing Granulation Drying Compaction Power Server Teacher GNN (PyTorch) ↓ Teaches Student Model (Knowledge Distillation) Models Transferring data to the server to be analyzed ESP32-P4 Edge Chip Student Flashed (C Headers) XGBoost Decision Trees ↻ Closed-Loop Inference OPC-UA / MQTT Control Commands (Adjusting Setpoints)
View Source on GitHub