predict.py
| 1 | """ |
| 2 | Inference Script for MHD Hybrid Nanofluid Thermal Management Model |
| 3 | |
| 4 | Provides easy-to-use interface for: |
| 5 | 1. Single-point prediction |
| 6 | 2. Batch prediction |
| 7 | 3. Parameter sweep analysis |
| 8 | 4. PSO optimization |
| 9 | """ |
| 10 | |
| 11 | import torch |
| 12 | import numpy as np |
| 13 | import json |
| 14 | import os |
| 15 | |
| 16 | from model import ThermalSurrogateModel, DataNormalizer, get_model_config |
| 17 | from pso_optimizer import PSOOptimizer |
| 18 | |
| 19 | |
| 20 | class ThermalPredictor: |
| 21 | """High-level prediction interface for the thermal management model.""" |
| 22 | |
| 23 | def __init__(self, model_dir='/app/outputs'): |
| 24 | """Load model, normalizer, and config.""" |
| 25 | self.config = get_model_config() |
| 26 | |
| 27 | # Load model |
| 28 | self.model = ThermalSurrogateModel( |
| 29 | input_dim=self.config['input_dim'], |
| 30 | hidden_dims=self.config['hidden_dims'], |
| 31 | output_dim=self.config['output_dim'], |
| 32 | dropout=0.0 |
| 33 | ) |
| 34 | self.model.load_state_dict(torch.load( |
| 35 | os.path.join(model_dir, 'model.pt'), |
| 36 | weights_only=True |
| 37 | )) |
| 38 | self.model.eval() |
| 39 | |
| 40 | # Load normalizer |
| 41 | self.normalizer = DataNormalizer.load( |
| 42 | os.path.join(model_dir, 'normalizer.json') |
| 43 | ) |
| 44 | |
| 45 | self.output_names = self.config['output_features'] |
| 46 | self.input_names = self.config['input_features'] |
| 47 | |
| 48 | def predict(self, Ha, phi, u_in): |
| 49 | """ |
| 50 | Predict thermal performance for given parameters. |
| 51 | |
| 52 | Args: |
| 53 | Ha: Hartmann number (0-60) |
| 54 | phi: Nanoparticle volume fraction (0.01-0.05) |
| 55 | u_in: Inlet flow velocity (0.05-0.30 m/s) |
| 56 | |
| 57 | Returns: |
| 58 | dict with predictions for all output variables |
| 59 | """ |
| 60 | X = np.array([[Ha, phi, u_in]], dtype=np.float32) |
| 61 | X_norm = self.normalizer.transform_input(X) |
| 62 | |
| 63 | with torch.no_grad(): |
| 64 | pred_norm = self.model(torch.tensor(X_norm, dtype=torch.float32)).cpu().numpy() |
| 65 | |
| 66 | pred = self.normalizer.inverse_transform_output(pred_norm)[0] |
| 67 | |
| 68 | result = {} |
| 69 | for i, name in enumerate(self.output_names): |
| 70 | result[name] = float(pred[i]) |
| 71 | |
| 72 | # Add derived metrics |
| 73 | result['thermal_runaway_risk'] = 'High' if result['T_max'] > 55 else \ |
| 74 | 'Moderate' if result['T_max'] > 45 else \ |
| 75 | 'Low' if result['T_max'] > 38 else 'Very Low' |
| 76 | result['within_safe_range'] = result['T_max'] <= 40.0 |
| 77 | |
| 78 | return result |
| 79 | |
| 80 | def batch_predict(self, params_list): |
| 81 | """ |
| 82 | Predict for multiple parameter sets. |
| 83 | |
| 84 | Args: |
| 85 | params_list: list of (Ha, phi, u_in) tuples |
| 86 | |
| 87 | Returns: |
| 88 | list of result dicts |
| 89 | """ |
| 90 | return [self.predict(*p) for p in params_list] |
| 91 | |
| 92 | def parameter_sweep(self, param_name, values, fixed_params): |
| 93 | """ |
| 94 | Sweep one parameter while holding others fixed. |
| 95 | |
| 96 | Args: |
| 97 | param_name: 'Ha', 'phi', or 'u_in' |
| 98 | values: array of values to sweep |
| 99 | fixed_params: dict with fixed values for other params |
| 100 | """ |
| 101 | results = [] |
| 102 | for val in values: |
| 103 | params = fixed_params.copy() |
| 104 | params[param_name] = val |
| 105 | pred = self.predict(params['Ha'], params['phi'], params['u_in']) |
| 106 | pred[param_name] = val |
| 107 | results.append(pred) |
| 108 | return results |
| 109 | |
| 110 | def optimize(self, alpha=0.55, n_particles=50, max_iter=200): |
| 111 | """ |
| 112 | Run PSO optimization. |
| 113 | |
| 114 | Args: |
| 115 | alpha: Weight for T_max (1-alpha for S_gen) |
| 116 | |
| 117 | Returns: |
| 118 | optimal_params, optimal_performance |
| 119 | """ |
| 120 | pso = PSOOptimizer( |
| 121 | model=self.model, |
| 122 | normalizer=self.normalizer, |
| 123 | n_particles=n_particles, |
| 124 | max_iter=max_iter, |
| 125 | alpha=alpha |
| 126 | ) |
| 127 | best_pos, best_pred, convergence = pso.optimize(verbose=False) |
| 128 | |
| 129 | params = { |
| 130 | 'Ha': float(best_pos[0]), |
| 131 | 'phi': float(best_pos[1]), |
| 132 | 'u_in': float(best_pos[2]), |
| 133 | } |
| 134 | |
| 135 | performance = {} |
| 136 | for i, name in enumerate(self.output_names): |
| 137 | performance[name] = float(best_pred[i]) |
| 138 | |
| 139 | return params, performance, convergence |
| 140 | |
| 141 | |
| 142 | def demo(): |
| 143 | """Run demonstration of the prediction pipeline.""" |
| 144 | print("=" * 65) |
| 145 | print(" MHD Hybrid Nanofluid EV Battery Thermal Management Model") |
| 146 | print(" Neural Network Surrogate with PSO Optimization") |
| 147 | print("=" * 65) |
| 148 | |
| 149 | predictor = ThermalPredictor() |
| 150 | |
| 151 | # 1. Single prediction at paper's PSO optimal |
| 152 | print("\n1. PREDICTION AT PAPER'S PSO OPTIMAL POINT") |
| 153 | print("-" * 50) |
| 154 | result = predictor.predict(Ha=32.4, phi=0.038, u_in=0.187) |
| 155 | print(f" Input: Ha=32.4, φ=0.038, u₀=0.187 m/s") |
| 156 | print(f" T_max: {result['T_max']:.2f}°C (Paper: 40.8°C)") |
| 157 | print(f" Nusselt (Nu): {result['Nu']:.2f} (Paper: 18.7)") |
| 158 | print(f" Entropy (S_gen):{result['S_gen']:.4f}") |
| 159 | print(f" Cell-to-Cell ΔT:{result['delta_T']:.2f}°C") |
| 160 | print(f" k_ratio: {result['k_ratio']:.3f}") |
| 161 | print(f" BL Suppression: {result['BL_suppression']:.2f}%") |
| 162 | print(f" Runaway Risk: {result['thermal_runaway_risk']}") |
| 163 | |
| 164 | # 2. Conventional cooling comparison |
| 165 | print("\n2. CONVENTIONAL COOLING (No MHD, Low Nanoparticle)") |
| 166 | print("-" * 50) |
| 167 | result_conv = predictor.predict(Ha=0, phi=0.01, u_in=0.15) |
| 168 | print(f" Input: Ha=0, φ=0.01, u₀=0.15 m/s") |
| 169 | print(f" T_max: {result_conv['T_max']:.2f}°C (Paper: ~61.3°C)") |
| 170 | print(f" Nusselt (Nu): {result_conv['Nu']:.2f} (Paper: ~12.4)") |
| 171 | print(f" Runaway Risk: {result_conv['thermal_runaway_risk']}") |
| 172 | |
| 173 | # 3. Improvement calculation |
| 174 | print("\n3. IMPROVEMENT: OPTIMIZED vs CONVENTIONAL") |
| 175 | print("-" * 50) |
| 176 | T_reduction = (result_conv['T_max'] - result['T_max']) / result_conv['T_max'] * 100 |
| 177 | Nu_improvement = (result['Nu'] - result_conv['Nu']) / result_conv['Nu'] * 100 |
| 178 | print(f" Temperature Reduction: {T_reduction:.1f}% (Paper: 33.4%)") |
| 179 | print(f" Nu Improvement: {Nu_improvement:.1f}% (Paper: 50.8%)") |
| 180 | |
| 181 | # 4. Parameter sweep: effect of Hartmann number |
| 182 | print("\n4. PARAMETER SWEEP: EFFECT OF HARTMANN NUMBER") |
| 183 | print("-" * 50) |
| 184 | print(f" {'Ha':>5} | {'T_max (°C)':>12} | {'Nu':>8} | {'BL Supp %':>10} | {'S_gen':>8}") |
| 185 | print(f" {'':->5}-+-{'':-<12}-+-{'':-<8}-+-{'':-<10}-+-{'':-<8}") |
| 186 | |
| 187 | Ha_sweep = predictor.parameter_sweep( |
| 188 | 'Ha', |
| 189 | np.arange(0, 61, 10), |
| 190 | fixed_params={'Ha': 0, 'phi': 0.03, 'u_in': 0.187} |
| 191 | ) |
| 192 | for r in Ha_sweep: |
| 193 | print(f" {r['Ha']:5.0f} | {r['T_max']:12.2f} | {r['Nu']:8.2f} | " |
| 194 | f"{r['BL_suppression']:10.2f} | {r['S_gen']:8.4f}") |
| 195 | |
| 196 | # 5. Run PSO optimization |
| 197 | print("\n5. PSO OPTIMIZATION (Finding optimal parameters)") |
| 198 | print("-" * 50) |
| 199 | optimal_params, optimal_perf, _ = predictor.optimize(alpha=0.55) |
| 200 | print(f" Optimal Ha: {optimal_params['Ha']:.2f}") |
| 201 | print(f" Optimal φ: {optimal_params['phi']:.4f}") |
| 202 | print(f" Optimal u₀: {optimal_params['u_in']:.4f} m/s") |
| 203 | print(f" → T_max: {optimal_perf['T_max']:.2f}°C") |
| 204 | print(f" → Nu: {optimal_perf['Nu']:.2f}") |
| 205 | print(f" → S_gen: {optimal_perf['S_gen']:.4f}") |
| 206 | |
| 207 | print("\n" + "=" * 65) |
| 208 | print(" Pipeline complete! Model ready for deployment.") |
| 209 | print("=" * 65) |
| 210 | |
| 211 | |
| 212 | if __name__ == "__main__": |
| 213 | demo() |
| 214 | |