comsol_opt

Python orchestration for a multi-scale COMSOL wafer-warpage calibration flow.

The active flow is configured as a layered DAG:

process step
  -> device_main / device_aux / onon_device RVEs
  -> mat1 / mat2 / mat3 / mat4 RVEs
  -> die RVE
  -> wafer model
  -> bow_x / bow_y comparison
  -> staged calibration

The Python layer owns config expansion, parameter TXT staging, state inheritance, backend dispatch, mock validation, loss calculation, and calibration. A real COMSOL worker should consume only step_input.json plus the parameter TXT files referenced there.

Quick Start

python3 -m unittest tests/test_workflow.py -v
python3 python/run_flow.py --backend dryrun --run-id dryrun_001
python3 python/run_flow.py --backend mock --run-id mock_001
python3 python/compute_loss.py --summary runs/mock_001/summary.csv
python3 python/calibration_optuna.py --backend mock --run-id calib_001 --n-trials 5

Run selected staged calibration:

python3 python/calibration_optuna.py \
  --backend mock \
  --mode staged \
  --stages G0_init,G2_trench_release \
  --run-id staged_001 \
  --n-trials 5

Active Layout

configs/
├── flow.yaml
├── templates.yaml
├── extractors.yaml
├── calibration_space.yaml
├── steps/
├── params/
└── experiments/

• configs/flow.yaml is the global entrypoint. It only lists enabled process steps and global config files.
• configs/steps/*.yaml defines each step DAG. Every node must explicitly declare action: run or action: inherit.
• configs/templates.yaml registers COMSOL template paths.
• configs/extractors.yaml registers study/evaluation tags.
• configs/params/*.txt stores COMSOL-style parameters. YAML is not used for the active runtime parameter source.
• configs/calibration_space.yaml stores optimizer bounds, priors, scales, and units.

Legacy compatibility files live under archive/legacy_config_scheme/. The active seed flow currently enables S00 and S02. configs/calibration_space.yaml can keep planned later stages, but default calibration only selects groups whose target step is enabled in configs/flow.yaml.

Parameter Contract

For each step, the orchestrator copies parameter TXT files into the step output directory and records:

• parameter_txt_paths: actual staged TXT file paths.
• parameter_txt_order: load order, normally global_params -> step params -> calibration_override.

When no explicit --params file is supplied, each layered DAG step uses parameters parsed from its own TXT stack. An explicit params YAML is treated as a trial/global override for calibration. Cache keys include staged TXT file content hashes, so COMSOL-only TXT changes invalidate cached step outputs.

TXT parsing rules:

1. Empty lines and # comments are ignored.
2. Each row is name expression description....
3. COMSOL expressions such as 100[MPa] and 120[nm] are preserved for the worker.
4. Later files override earlier files with the same parameter name.

DAG State Contract

Each state_out.json contains:

• rve.device_main
• rve.device_aux
• rve.onon_device
• rve.mat1 through rve.mat4
• rve.die
• wafer_result
• materials_state
• geometry_state
• history

The wafer node consumes only rve.die and rve.onon_device. The die node consumes only mat1 through mat4. Inherited nodes must already exist in state_in and have valid=true.

The Python validator checks DAG node actions, types, outputs, required inputs, template resolution, and topological RVE references before dispatching a backend.

Outputs

Each mock or COMSOL step should write:

step_input.json
state_out.json
step_result.json
manifest.json
*_rve.json
wafer_result.json

Global calibration cache lives under runs/<run-id>/.cache. Staged calibration cache lives under runs/<run-id>/out/<step>_<group>/.cache.

COMSOL Worker

java/ComsolStepWorker.java follows the active nodes contract: it loads TXT parameters in parameter_txt_order, runs or inherits DAG nodes, writes configured node result files, writes manifest.json, preserves material/geometry/history state, and outputs the same high-level JSON shape as the mock backend.

The worker still uses provisional RVE transfer assumptions. It injects upstream values as input_<slot>_sxx, input_<slot>_syy, input_<slot>_rho, input_<slot>_d11, and input_<slot>_d22. Validate or replace these names against the real COMSOL templates before production calibration.