ASD Agent Tutorial

Reproducible Python software for studying AI-agent loops, Bayesian active learning, constrained multi-objective Bayesian optimization, and manual laboratory handoff in an educational area-selective deposition (ASD) simulator.

This project does not predict real HfO2/MoS2 chemistry. Its equations and scenarios are teaching models for software design, optimization, benchmarking, and failure analysis. Do not use its recommendations as laboratory process guidance.

Features

• Seeded two-surface virtual laboratory with GA and NGA growth, nucleation delays, saturating doses, inhibitor blocking, temperature response, noise, and process time.
• Legacy random search, grid search, deterministic rule-based agent, and optional OpenAI Responses API agent.
• Stage 1 fixed-grid, generic-GP, and physics-informed-GP saturation learning.
• Stage 2 constrained noisy multi-objective BO over precursor dose, temperature, and integer cycle count.
• Hybrid LLM-BO orchestration with immutable candidate IDs and a fake LLM by default.
• Paired research profiles, bootstrap confidence intervals, McNemar comparisons, Wilcoxon tests, Holm correction, and publication-table exports.
• Human-operated laboratory plan export and validated measurement ingestion. No autonomous hardware control is included.
• Eight executable tutorial notebooks and deterministic plotting scripts.

Requirements

• Python 3.12 through 3.14.
• The full suite is currently verified on Python 3.12.13.
• CPU execution is supported. A GPU is optional and not required by the smoke profiles.

Python 3.13 and 3.14 are compatibility targets, not verified environments in the current repository record.

Installation

Create a virtual environment and install the package in editable mode:

py -3.12 -m venv .venv
.venv\Scripts\python -m pip install --upgrade pip
.venv\Scripts\python -m pip install -e ".[dev,bo,bo-gp,bo-analysis,notebooks]"

Optional extras:

Extra	Purpose
bo	NumPy/SciPy analysis helpers
bo-gp	PyTorch, GPyTorch, and BoTorch models
bo-ax	Optional Ax experiments; not required by the implemented runners
bo-analysis	Statistical and plotting utilities
llm	Optional live OpenAI Responses API paths
notebooks	JupyterLab and notebook file support
dev	pytest, Ruff, mypy, and type stubs

Quick Start

Run the no-API legacy tutorial:

asd-demo --scenario inherent_selectivity --agent rule_based
asd-benchmark --repetitions 20 --output-dir results/legacy_benchmark

Run Bayesian-optimization workflows through the integrated CLI:

asd-agent stage1 run --profile smoke --scenario fast_mono --method physics_gp `
  --output-dir results/stage1_smoke

asd-agent stage2 run --profile smoke --scenario inherent_selectivity `
  --method stage2_mobo --output-dir results/stage2_smoke

asd-agent hybrid run --scenario narrow_selective_window `
  --mode hybrid_advisory --fake-llm --output-dir results/hybrid_smoke

The fake LLM is deterministic and makes no network calls.

CLI

asd-agent stage1 run|compare
asd-agent stage2 run|compare
asd-agent hybrid run
asd-agent study run|analyze
asd-agent lab export|ingest

Show command-specific options with asd-agent <group> <command> --help. Seeds, budgets, scenarios, profiles, and output directories are explicit command arguments or validated YAML values.

Research studies

asd-agent study run --config smoke --output-dir results/study_smoke
asd-agent study analyze --input results/study_smoke/research_results.json `
  --output results/study_smoke/analysis

Larger profiles are configured but are never run automatically:

python scripts/run_stage1_study.py --profile pilot --output-dir results/stage1_pilot
python scripts/run_stage2_benchmark.py --profile pilot --output-dir results/stage2_pilot
python scripts/run_research_study.py --profile paper_non_llm `
  --output-dir results/paper_non_llm
python scripts/run_research_study.py --profile paper_llm `
  --output-dir results/paper_llm

Review compute budgets, resolved dependencies, and the research protocol before starting pilot or paper-scale profiles.

Optional live LLM

Live calls require the llm extra, environment credentials, a model name, and an explicit live flag:

$env:OPENAI_API_KEY = "..."
$env:OPENAI_MODEL = "your-responses-api-model"
asd-agent hybrid run --scenario narrow_selective_window `
  --mode hybrid_advisory --live-llm --output-dir results/hybrid_live

Tests and default smoke commands never make live API calls. The live hybrid adapter is import- and schema-tested but has not been exercised against the API in this repository.

Manual Laboratory Handoff

Export one validated, pending experiment plan:

asd-agent lab export --scenario inherent_selectivity --run-id manual_001 `
  --output results/manual_001

After a human operator completes and reviews the measurement template, import it in a later process:

asd-agent lab ingest --scenario inherent_selectivity `
  --plan results/manual_001/manual_lab_plan.json `
  --measurements results/manual_001/completed_measurements.csv `
  --output results/manual_001/optimizer_observations.json

Imported measurements require GA/NGA thickness in nanometers, a measurement method, replicate ID, quality-control status, and operator notes. See docs/lab_validation_protocol.md.

Configuration

YAML files in configs/ define:

• the three legacy ASD scenarios;
• six Stage 1 saturation-process families;
• eight Stage 2 constrained optimization scenarios;
• Stage 1 and Stage 2 smoke/pilot profiles;
• paired smoke, pilot, paper non-LLM, and paper LLM study profiles.

Hidden simulator parameters remain in the backend configuration. Optimizers and LLM contexts receive explicit optimizer-facing views that exclude oracle values and hidden process constants.

Outputs and Reproducibility

Generated files are written below results/ and ignored by Git except for results/.gitkeep. Depending on the command, outputs include:

• experiment ledgers and optimizer proposals in JSON/CSV;
• run manifests with Git commit, configuration hash, Python/dependency versions, named seeds, method settings, and timestamps;
• hypervolume trajectories, failure categories, and constraint outcomes;
• CSV source data beside PNG, SVG, and PDF figures;
• JSON, CSV, Markdown, and LaTeX statistical tables;
• manual laboratory plans, templates, and imported observations.

Every stochastic simulator and optimizer path accepts an explicit seed. Paired studies persist separate simulator, measurement-noise, initialization, BO, and LLM seed streams.

Repository Layout

configs/             Scenario and study YAML
docs/                Protocols, implementation records, and model cards
notebooks/           Eight tutorial notebooks
prompts/             Prompt-data guidance; no agent traces
scripts/             Reproducible smoke, benchmark, and plotting entry points
src/asd_agent/       Installable package
src/asd_agent/bo/    BO, hybrid, statistics, oracle, and lab-handoff modules
tests/               Unit and integration tests
results/             Ignored generated outputs

Development

Run the complete local verification suite:

ruff format --check .
ruff check .
mypy src
pytest

The test suite covers simulator behavior, safety bounds, schemas, oracle isolation, tested recommendations, GP posteriors, mixed-variable MOBO, hybrid bypass prevention, paired statistics, save/restore, and manual-lab ingestion.

Documentation

• Research protocol
• Tutorial outline
• Laboratory validation protocol
• Generic GP model card
• Physics-informed GP model card
• Constrained MOBO model card
• Hybrid agent model card

Limitations

• The simulator is not calibrated to experimental chemistry.
• Oracle fronts and true saturation values exist only for retrospective virtual evaluation and are not laboratory ground truth.
• Independent Stage 2 outcome GPs do not exactly preserve the algebraic relationship between GA, NGA, and selectivity; consistency diagnostics are recorded.
• Smoke and pilot profiles verify software pipelines but are not research conclusions.
• Python 3.13/3.14, GPU execution, Ax integration, live LLM behavior, and paper-scale runtime have not been verified in the current environment.
• Manual plan validation is not a substitute for institutional safety review, equipment limits, calibration, replicates, uncertainty analysis, or qualified human approval.

Scientific 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,” Journal of Vacuum Science & Technology A 44, 043410 (2026), DOI: 10.1116/6.0005313.

The implementation, equations, parameter values, scenarios, and benchmark claims in this repository are independent educational constructs.