NEMESIS

Initialize a clean run directory

The script sets WORK_DIR, derives paths for the DUT, netlist, specs, prompts, models, and reports, then removes any existing work directory with the same name. It recreates subdirectories for testbenches, netlist, prompts, llm_outputs, ota_models, ota_model_performance, spice_performance, and specs.

It copies the universal testbenches from Universal/testbenches/ and the selected DUT netlist from Universal/dut/<WORK_DIR>.sp.

Generate a simplified netlist for LLM context

generate_netlist_for_llm.py extracts the DUT subcircuit and compacts transistor lines into a cleaner representation. This reduces prompt noise while preserving topology, device names, terminals, model types, and sizing parameters.

Create design_params.sp

generate_design_params.py scans the DUT devices and creates a SPICE parameter template containing width and length variables. The default script call sets device length to 180, and later sizing stages update the width values.

Generate an operating-point HSPICE deck

generate_op_hspice.py combines the selected DUT and tb_op.sp template into tb_op_temp.sp. The temporary deck is formatted for reliable operating-point table extraction by Python.

Build the gm/Id estimator prompt

generate_gmid_estimator_agent_prompt.py prepares an LLM prompt using the simplified netlist plus design context: Vdd, Vss, ibias, channel length, application class, and design intent.

Run the LLM to generate specs

call_codex.py sends the gm/Id prompt and topology figure to the selected model. The output is written to <WORK_DIR>/specs/<WORK_DIR>.json and becomes the target file used by the sizer.

Generate the first equation-model prompt

generate_stage1_prompt.py creates prompt_stage1.txt from the simplified netlist. This prompt asks the LLM to produce structured performance equations for the selected OTA topology.

Generate candidate equations

The LLM consumes the current prompt and circuit figure, then writes llm_output_stageN.json. If this JSON is missing, the same iteration is retried without incrementing the stage counter.

Update equation history

extract_equations.py appends the generated equations into equation_history.json. This gives later feedback prompts memory of what has already been attempted.

Compile JSON equations into Python

compile_ota_model_script.py converts the structured JSON response into ota_model_stageN.py, an executable scalar evaluator used by the sizing and verification scripts.

Run gm/Id LUT-based sizing

design_sizer.py evaluates the compiled model against the generated specs and writes ota_model_estimation_stageN.1.json. With --optimize, it searches for operating points that satisfy gm/Id and saturation constraints.

If the model crashes or fails to size, repair_model.py sends the sizer error log, compiled model, and optional topology image back through the LLM. The repaired JSON is recompiled and retried up to MAX_MODEL_REPAIR_ATTEMPTS.

Write sized parameters back to SPICE

update_design_params.py maps the gm/Id sizing result into design_params_updated.sp. This file drives the HSPICE testbenches with model-derived widths and lengths.

Run HSPICE characterization

perform_hspice_simulations.py runs the universal test plan with the selected DUT, updated parameters, DUT wrapper, and local run directory. It produces a JSON report and appends sweep data to design_sweep_log.csv.

Extract transistor operating points

extract_op.py parses run_op.lis into op_results.json. These simulator operating-point values are then supplied to the model evaluator for a more direct comparison against SPICE.

Re-run sizing in verification mode

design_sizer.py --verify --opjson evaluates the same compiled model using extracted HSPICE operating-point data. The result is written as ota_model_estimation_stageN.2.json. If verification output is missing, the iteration is discarded and regenerated.

Analyze mismatch and create the next prompt

feedback_manager.py compares HSPICE performance with equation-model performance using a default 15.0 percent threshold. If all equations are accurate, convergence is declared. Otherwise the script writes prompt_stageN+1.txt with targeted feedback and continues.

Generate a vectorized final evaluator

After convergence, vectorization_agent.py converts the scalar evaluator into ota_model_stageN_vectorized.py. The vectorized script is intended for faster batched evaluation once the equation set has already passed SPICE closure.