Inspiration
Inspired by studies showing that ~500 moderate-quality data points can yield >95% bidirectional prediction accuracy—proving that AI can break the “big data dependency” and make intelligent manufacturing practical. At the same time, LIG parameter optimization today relies heavily on manual trial‑and‑error, which is inefficient and poorly reproducible. This motivated us to build an end‑to‑end platform that maps target performance → optimal laser parameters.
What it does
The platform takes user‑specified substrate material and target properties (sheet resistance, EMI shielding effectiveness, frequency response, etc.) and automatically recommends laser power, frequency, pulse width, scanning speed, and other processing parameters. It supports both forward prediction (parameters → properties) and inverse design (properties → parameters), provides explainable recommendations, and continuously improves through experimental feedback.
How we built it
We built a bidirectional neural network (inverse design network + forward validation network) as the core, combined with Bayesian optimization for efficient exploration, and constructed a process knowledge graph for interpretability. The backend uses Python + FastAPI, the frontend React, and databases PostgreSQL and Neo4j, all in a modular microservices architecture.
Challenges we ran into
- Scarce & noisy data – adopted “black‑box deep learning” strategies with moderate‑quality small samples, plus data augmentation and transfer learning.
- Non‑unique inverse solutions – used a tandem network with the forward network as a regularizer to ensure physically feasible outputs.
- Generalization across substrates/lasers – introduced hierarchical embeddings and knowledge graphs to transfer process relationships.
- Explainability – retrieved similar cases and sensitivity analyses from the knowledge graph to build user trust.
Accomplishments that we're proud of
- Achieved ~95% prediction accuracy with limited data, matching state‑of‑the‑art literature.
- Built a closed‑loop bidirectional system where recommended parameters are validated by forward prediction with controlled deviation.
- Successfully applied the platform to multi‑objective scenarios like frequency‑selective surfaces (FSS).
- Established a continuous learning loop that evolves the model with each new experiment.
What we learned
- Data quality > quantity – careful experimental design that covers the parameter space matters most.
- Joint bidirectional training effectively resolves the ill‑posed inverse problem.
- AI can uncover parameter combinations beyond human biases, yielding up to 2× improvement in optimization.
- Cross‑disciplinary collaboration (AI + laser + materials) is the key to success.
What's next for AI‑Driven LIG Intelligent Processing Parameter Optimization Platform
- Short‑term: Expand substrate library, interface with real laser systems for closed‑loop control, and enhance multi‑objective optimization.
- Medium‑term: Build a digital twin simulation, establish an open data‑sharing community, and extend to other laser processing domains.
- Long‑term: Combine with automated robotic experiments for fully autonomous parameter discovery, and scale to industrial deployment as an intelligent manufacturing infrastructure.
Built With
- ai
- aioptimization
- laserprocessing
- smartmanufacturing
Log in or sign up for Devpost to join the conversation.