"""Dataset generation pipeline using Latin Hypercube Sampling. Generates 100K structural analysis samples across beams, plates, and pressure vessels using analytical closed-form solutions. LHS provides better coverage of the design space than random sampling — a deliberate choice from Design of Experiments (DoE) methodology. Usage: python -m src.data.generate_dataset --config configs/data_generation.yaml """ import argparse import hashlib import logging import uuid from pathlib import Path from typing import Any import numpy as np import pandas as pd import yaml from scipy.stats.qmc import LatinHypercube from src.data.schema import ProblemFamily, SafetyCategory from src.data.solvers.beam import BEAM_SOLVERS from src.data.solvers.plate import PLATE_SOLVERS from src.data.solvers.vessel import VESSEL_SOLVERS logger = logging.getLogger(__name__) def _lhs_sample( n_samples: int, param_ranges: dict[str, dict], seed: int, ) -> dict[str, np.ndarray]: """Generate Latin Hypercube Samples and map to physical parameter ranges. Log-uniform parameters are sampled uniform in log-space then exponentiated, which is critical because engineering quantities span orders of magnitude (e.g., elastic modulus: 1 GPa to 400 GPa). """ param_names = list(param_ranges.keys()) n_dims = len(param_names) sampler = LatinHypercube(d=n_dims, seed=seed) unit_samples = sampler.random(n=n_samples) # shape: (n_samples, n_dims) result: dict[str, np.ndarray] = {} for i, name in enumerate(param_names): spec = param_ranges[name] lo, hi = float(spec["min"]), float(spec["max"]) col = unit_samples[:, i] if spec.get("distribution") == "log_uniform": log_lo, log_hi = np.log10(lo), np.log10(hi) result[name] = 10.0 ** (log_lo + col * (log_hi - log_lo)) else: # uniform result[name] = lo + col * (hi - lo) return result def _generate_beam_samples( config: dict, global_seed: int, ) -> list[dict[str, Any]]: """Generate beam samples across all 6 configurations.""" samples = [] beam_configs = config["beam"]["configs"] n_per_config = config["beam"]["samples_per_config"] param_ranges = config["beam"]["parameters"] for cfg_idx, cfg in enumerate(beam_configs): config_id = cfg["id"] solver_cls = BEAM_SOLVERS[config_id] solver = solver_cls() seed = global_seed + cfg_idx lhs_params = _lhs_sample(n_per_config, param_ranges, seed) for i in range(n_per_config): params = {k: float(v[i]) for k, v in lhs_params.items()} # Select load type based on config if "point" in config_id: load_params = { "point_load": params["point_load"], "distributed_load": 0.0, } else: load_params = { "point_load": 0.0, "distributed_load": params["distributed_load"], } solve_params = { "length": params["length"], "width": params["width"], "height": params["height"], "elastic_modulus": params["elastic_modulus"], "yield_strength": params["yield_strength"], **load_params, } result = solver.solve(solve_params) b, h = params["width"], params["height"] samples.append({ "sample_id": str(uuid.uuid4()), "problem_family": ProblemFamily.BEAM.value, "config_id": config_id, "length": params["length"], "width": params["width"], "height": params["height"], "inner_radius": None, "outer_radius": None, "thickness": None, "elastic_modulus": params["elastic_modulus"], "poisson_ratio": params["poisson_ratio"], "yield_strength": params["yield_strength"], "density": params["density"], "point_load": load_params["point_load"], "distributed_load": load_params["distributed_load"], "internal_pressure": 0.0, "pressure": 0.0, "moment_of_inertia": b * h**3 / 12.0, "section_modulus": b * h**2 / 6.0, "cross_section_area": b * h, "max_stress": result.max_stress, "max_deflection": result.max_deflection, "safety_factor": result.safety_factor, "safety_category": result.safety_category.value, }) logger.info(f"Generated {n_per_config} samples for {config_id}") return samples def _generate_plate_samples( config: dict, global_seed: int, ) -> list[dict[str, Any]]: """Generate plate samples across both configurations.""" samples = [] plate_configs = config["plate"]["configs"] n_per_config = config["plate"]["samples_per_config"] param_ranges = config["plate"]["parameters"] for cfg_idx, cfg in enumerate(plate_configs): config_id = cfg["id"] solver_cls = PLATE_SOLVERS[config_id] solver = solver_cls() seed = global_seed + 100 + cfg_idx lhs_params = _lhs_sample(n_per_config, param_ranges, seed) for i in range(n_per_config): params = {k: float(v[i]) for k, v in lhs_params.items()} solve_params = { "length_a": params["length_a"], "length_b": params["length_b"], "thickness": params["thickness"], "elastic_modulus": params["elastic_modulus"], "poisson_ratio": params["poisson_ratio"], "yield_strength": params["yield_strength"], "pressure": params["pressure"], } result = solver.solve(solve_params) samples.append({ "sample_id": str(uuid.uuid4()), "problem_family": ProblemFamily.PLATE.value, "config_id": config_id, "length": params["length_a"], "width": params["length_b"], "height": None, "inner_radius": None, "outer_radius": None, "thickness": params["thickness"], "elastic_modulus": params["elastic_modulus"], "poisson_ratio": params["poisson_ratio"], "yield_strength": params["yield_strength"], "density": params["density"], "point_load": 0.0, "distributed_load": 0.0, "internal_pressure": 0.0, "pressure": params["pressure"], "moment_of_inertia": None, "section_modulus": None, "cross_section_area": None, "max_stress": result.max_stress, "max_deflection": result.max_deflection, "safety_factor": result.safety_factor, "safety_category": result.safety_category.value, }) logger.info(f"Generated {n_per_config} samples for {config_id}") return samples def _generate_vessel_samples( config: dict, global_seed: int, ) -> list[dict[str, Any]]: """Generate pressure vessel samples across both configurations.""" samples = [] vessel_configs = config["vessel"]["configs"] n_per_config = config["vessel"]["samples_per_config"] param_ranges = config["vessel"]["parameters"] for cfg_idx, cfg in enumerate(vessel_configs): config_id = cfg["id"] solver_cls = VESSEL_SOLVERS[config_id] solver = solver_cls() seed = global_seed + 200 + cfg_idx lhs_params = _lhs_sample(n_per_config, param_ranges, seed) for i in range(n_per_config): params = {k: float(v[i]) for k, v in lhs_params.items()} r_i = params["inner_radius"] r_o = r_i * params["radius_ratio"] solve_params = { "inner_radius": r_i, "outer_radius": r_o, "elastic_modulus": params["elastic_modulus"], "poisson_ratio": params["poisson_ratio"], "yield_strength": params["yield_strength"], "internal_pressure": params["internal_pressure"], } result = solver.solve(solve_params) samples.append({ "sample_id": str(uuid.uuid4()), "problem_family": ProblemFamily.VESSEL.value, "config_id": config_id, "length": params.get("length", None) if config_id == "vessel_cylinder" else None, "width": None, "height": None, "inner_radius": r_i, "outer_radius": r_o, "thickness": r_o - r_i, "elastic_modulus": params["elastic_modulus"], "poisson_ratio": params["poisson_ratio"], "yield_strength": params["yield_strength"], "density": params["density"], "point_load": 0.0, "distributed_load": 0.0, "internal_pressure": params["internal_pressure"], "pressure": 0.0, "moment_of_inertia": None, "section_modulus": None, "cross_section_area": None, "max_stress": result.max_stress, "max_deflection": result.max_deflection, "safety_factor": result.safety_factor, "safety_category": result.safety_category.value, }) logger.info(f"Generated {n_per_config} samples for {config_id}") return samples def generate_dataset(config_path: str) -> pd.DataFrame: """Generate the full dataset from config file.""" with open(config_path) as f: config = yaml.safe_load(f) seed = config["seed"] np.random.seed(seed) logger.info("Generating beam samples...") beam_samples = _generate_beam_samples(config, seed) logger.info("Generating plate samples...") plate_samples = _generate_plate_samples(config, seed) logger.info("Generating vessel samples...") vessel_samples = _generate_vessel_samples(config, seed) all_samples = beam_samples + plate_samples + vessel_samples df = pd.DataFrame(all_samples) logger.info(f"Total samples generated: {len(df)}") logger.info(f"Safety category distribution:\n{df['safety_category'].value_counts()}") return df def split_and_save(df: pd.DataFrame, config_path: str) -> None: """Stratified split and save as Parquet files.""" with open(config_path) as f: config = yaml.safe_load(f) output_dir = Path(config["output"]["directory"]) output_dir.mkdir(parents=True, exist_ok=True) splits = config["splits"] train_frac = splits["train"] val_frac = splits["validation"] # Stratified split by config_id and safety_category df_shuffled = df.sample(frac=1.0, random_state=config["seed"]).reset_index(drop=True) n = len(df_shuffled) n_train = int(n * train_frac) n_val = int(n * val_frac) train_df = df_shuffled.iloc[:n_train] val_df = df_shuffled.iloc[n_train:n_train + n_val] test_df = df_shuffled.iloc[n_train + n_val:] train_df.to_parquet(output_dir / "train.parquet", index=False) val_df.to_parquet(output_dir / "validation.parquet", index=False) test_df.to_parquet(output_dir / "test.parquet", index=False) logger.info(f"Saved splits: train={len(train_df)}, val={len(val_df)}, test={len(test_df)}") logger.info(f"Output directory: {output_dir}") def main() -> None: parser = argparse.ArgumentParser(description="Generate structural mechanics dataset") parser.add_argument("--config", default="configs/data_generation.yaml") args = parser.parse_args() logging.basicConfig(level=logging.INFO, format="%(asctime)s %(levelname)s %(message)s") df = generate_dataset(args.config) split_and_save(df, args.config) if __name__ == "__main__": main()