TWLab/femtosecond-laser-hydrogel-etching-model

Femtosecond Laser Hydrogel Etching - RL-Optimized Model v2

🔬 Overview

An ML system combining a surrogate prediction model (XGBoost + Neural Network ensemble) with a Reinforcement Learning optimizer (SAC) for femtosecond laser hydrogel etching parameter optimization.

Focus materials: GelMA and HAMA (65% of 12,000 training samples)

Architecture

┌──────────────────────────────────────────────────────────────────┐
│  SAC RL Agent (Parameter Optimizer) - 100% Success Rate          │
│  • Observes: current params + predicted geometry + target        │
│  • Actions: parameter deltas (±15% of range per step)            │
│  • Reward: improvement + proximity + goal bonus                  │
│  • Converges in ~4 steps to target geometry                      │
├──────────────────────────────────────────────────────────────────┤
│  XGBoost + NN Ensemble (Surrogate Environment)                   │
│  • XGBoost: 600 trees, depth 7, per-target                      │
│  • Neural Network: [128, 64, 32], BatchNorm, Dropout             │
│  • Weighted ensemble: 60% XGB + 40% NN                          │
└──────────────────────────────────────────────────────────────────┘

Performance

Surrogate Model (Geometry Prediction)

Target	R²	MAE
etch_depth_um	0.9745	1.05 µm
etch_width_um	0.9845	2.33 µm
surface_roughness_Sa_um	0.8409	28.18
aspect_ratio	0.7074	0.49
side_wall_angle_deg	0.9326	4.03°

RL Optimizer (SAC v2 - Depth + Width Focus)

Metric	Value
Success rate	100%
Avg relative error	0.107
Avg absolute error	2.60 µm
Convergence speed	~4 steps
Algorithm	SAC (γ=0.7, lr=1e-3)
Training timesteps	20,000

Design Principles (from literature)

Based on published RL-for-laser research:

• SAC dominates PPO/TD3 for continuous laser control (TempoRL, arxiv:2304.12187)
• Delta actions (±15% of param range/step) for machine safety (arxiv:2503.00499)
• γ = 0.7 optimal for episodic manufacturing control (not 0.99)
• Potential-based reward shaping with improvement signal + proximity bonus
• Domain randomization: material + target randomized each episode

Supported Materials (18 variants)

GelMA (40% of data)

Variant	Water Content	Threshold Fluence	Young's Modulus
GelMA 5%	92%	1.8 J/cm²	3.5 kPa
GelMA 7%	89%	1.4 J/cm²	8 kPa
GelMA 10%	85%	1.0 J/cm²	18 kPa
GelMA 15%	80%	0.7 J/cm²	45 kPa
GelMA 20%	75%	0.5 J/cm²	90 kPa

HAMA (25% of data)

Variant	Water Content	Threshold Fluence	Young's Modulus
HAMA 1%	96%	2.5 J/cm²	1.2 kPa
HAMA 2%	94%	2.0 J/cm²	3.5 kPa
HAMA 3%	92%	1.6 J/cm²	8 kPa
HAMA 5%	88%	1.2 J/cm²	20 kPa

GelMA-HAMA Composites (20% of data)

GelMA5+HAMA1, GelMA7+HAMA2, GelMA10+HAMA3, GelMA15+HAMA5

Other Materials (15% of data)

PEG, PEGDA, Collagen, Alginate, Silk fibroin

Input Features (21 total)

Laser Parameters (8)

power_mW, repetition_rate_kHz, scan_speed_mm_s, pulse_duration_fs, wavelength_nm, num_passes, spot_diameter_um, focal_offset_um

Material Properties (9)

water_content, threshold_fluence_J_cm2, absorption_depth_nm, incubation_coefficient, refractive_index, youngs_modulus_kPa, degree_of_methacrylation, crosslink_density_mol_m3, two_photon_cross_section_GM

Derived Physics Features (4)

pulse_energy_uJ, peak_fluence_J_cm2, effective_pulses, overlap_percent

Usage

Predict etching geometry:

import xgboost as xgb, torch, joblib, numpy as np

# Load models
scaler_X = joblib.load("scaler_X.joblib")
scaler_y = joblib.load("scaler_y.joblib")
xgb_models = {t: xgb.XGBRegressor() for t in targets}
for t in targets: xgb_models[t].load_model(f"xgb_{t}.json")

# Predict
features = [200, 1000, 1.0, 200, 800, 5, 10, 0, ...]  # 21 features
xgb_pred = np.array([xgb_models[t].predict([features])[0] for t in targets])

Optimize parameters with RL:

from stable_baselines3 import SAC

sac = SAC.load("sac_optimizer_v2")
# Set target: depth=10µm, width=20µm
# Agent finds optimal: power, rep_rate, speed, passes, spot_diameter
obs = env.reset(target=[10, 20])
for step in range(20):
    action, _ = sac.predict(obs, deterministic=True)
    obs, reward, done, _, info = env.step(action)
    if done: break
# Optimized params available in env.params

Dataset

12,000 samples with physics-informed synthetic data grounded in:

• Beer-Lambert ablation model with two-photon absorption
• Gaussian beam width model
• Material-specific incubation effects
• Literature-validated parameter ranges

Available at: TWLAb/femtosecond-laser-hydrogel-etching-data

Scientific References

1. Petit et al., "AI-Supported Prediction of Femtosecond Laser Micromachining Parameters", JLMN 2025
2. "TempoRL: Laser Pulse Temporal Shape Optimization with DRL", arxiv:2304.12187
3. Capuano et al., "Shaping Laser Pulses with RL", arxiv:2503.00499
4. "Optimizing Operation Recipes with RL", arxiv:2511.16297
5. Femtosecond laser induced densification within cell-laden hydrogels, Biofabrication 2019
6. Femtosecond Laser Densification of Hydrogels, ACS Appl. Mater. Interfaces 2022
7. Harnessing the potential of HAMA hydrogel, Bioactive Materials 2024
8. Behbahani et al., "ML-driven laser machining of alumina ceramics", Physica Scripta 2023

Generated by ML Intern

This model repository was generated by ML Intern, an agent for machine learning research and development on the Hugging Face Hub.

• Try ML Intern: https://smolagents-ml-intern.hf.space
• Source code: https://github.com/huggingface/ml-intern