ASD Agent Tutorial

Educational Python tutorial for an AI-agent loop in a virtual area-selective deposition (ASD) laboratory. The project is inspired by a 2026 ALD optimization paper in which reasoning-model agents iteratively propose dose times, receive measurements, and decide whether a process is saturated enough to recommend or abandon.

This repository does not predict real HfO2/MoS2 chemistry. It uses a toy simulator to teach agent design, optimization, benchmarking, reproducibility, and failure modes.

Python Target

The tutorial targets Python 3.14+, following the latest stable CPython release line available from python.org on July 10, 2026: Python 3.14.6. Python 3.15 is still pre-release at that date.

Quick Start

py -3.14 -m venv .venv
.venv\Scripts\python -m pip install -U pip
.venv\Scripts\python -m pip install -e ".[dev,llm,notebooks]"

Run a no-API deterministic demo:

asd-demo --scenario inherent_selectivity --agent rule_based

Run the benchmark with at least 20 repetitions:

asd-benchmark --repetitions 20 --output-dir results/benchmark_demo

Use the LLM agent:

set OPENAI_API_KEY=sk-...
set OPENAI_MODEL=gpt-5.6
asd-demo --scenario inhibitor_selectivity --agent llm

The LLM path uses the OpenAI Responses API with strict function tools: propose_experiments and finish_optimization.

What Is Modeled

• Growth area (GA) and non-growth area (NGA, representing MoS2 in the tutorial story).
• Saturating precursor-dose and coreactant-dose responses.
• Surface-specific maximum growth per cycle.
• Surface-specific nucleation delay.
• Optional inhibitor blocking, stronger on NGA in the inhibitor scenario.
• Optional temperature response.
• Configurable Gaussian measurement noise.
• Thickness after N cycles.
• Selectivity (GA - NGA) / (GA + NGA), with zero returned when both are zero.

Scenarios

• inherent_selectivity: GA nucleates quickly; NGA has a long nucleation delay.
• inhibitor_selectivity: inhibitor strongly blocks NGA and weakly affects GA.
• impossible_selectivity: tested bounds contain no meaningful selective window.

Default Success Criteria

• GA thickness >= 5 nm.
• NGA thickness <= 0.5 nm.
• Selectivity >= 0.80.
• Temperature and dose times inside configured safety bounds.

Repository Layout

configs/        Scenario definitions
src/asd_agent/  Simulator, agents, baselines, benchmark, plotting
notebooks/      Four tutorial notebooks
tests/          Pytest suite
results/        Generated ledgers, metadata, plots, and benchmark outputs

Reproducibility

Every run records configuration, timestamp, seed, model name, token usage, and an experiment ledger as CSV and JSON. Random-search and simulator noise use explicit NumPy seeds. The default demo mode uses the deterministic rule-based agent and does not require API credentials.

Source Inspiration

The high-level agent loop is inspired by Angel Yanguas-Gil, "Performance of AI agents based on reasoning language models on ALD process optimization tasks," J. Vac. Sci. Technol. A 44, 043410 (2026), DOI: 10.1116/6.0005313. The simulator and scenarios here are intentionally simplified and educational.