PALTO: Physics-Informed Active Learning for Tri-Gate FinFET Design Optimization for Vertical Power Delivery

Researchers demonstrate a physics-informed machine learning framework called PALTO for optimizing GaN tri-gate FinFET designs in power delivery systems, achieving 2× better performance than industrial benchmarks through intelligent exploration of device parameters. The approach addresses computational limitations of traditional TCAD simulations while enabling discovery of optimal gate-to-drain configurations and channel thickness ratios.

This research tackles a fundamental challenge in semiconductor device engineering: the computational intractability of optimizing complex power transistor designs. Traditional TCAD (Technology Computer-Aided Design) simulations require exhaustive parameter sweeps that consume substantial computing resources while often failing to locate truly optimal configurations in high-dimensional design spaces. By integrating physics-informed principles with active learning algorithms, the PALTO framework intelligently prioritizes which simulations to run, dramatically reducing computational overhead while maintaining accuracy.

The work emerges from decades of debate around optimal GaN-to-AlGaN thickness ratios in heterojunction devices, a design question that has lacked systematic resolution. The framework's ability to navigate this question systematically represents a meaningful advance in device engineering methodology. The identified designs, particularly device D1 with its 300-fin configuration delivering 3.3A at 0.49 ohms on-resistance, demonstrate practical improvements over existing industrial solutions.

For the semiconductor and power electronics industry, this research signals a shift toward ML-augmented design workflows. Rather than replacing physics-based simulation, the approach leverages machine learning to guide expensive computations more intelligently. This hybrid methodology could accelerate development cycles for application-specific power devices, particularly relevant as demand grows for efficient power delivery in data centers, automotive systems, and renewable energy infrastructure.

Future implementations may extend this framework to other device topologies and materials systems, potentially establishing active learning as standard practice in advanced semiconductor design. The demonstrated 2× efficiency improvement over benchmarks suggests meaningful commercial value if translated into manufacturing processes.

• →Physics-informed active learning framework reduces computational burden of semiconductor device optimization while improving design quality.
• →Optimized GaN FinFET device achieves 2× better performance metrics than industrial benchmarks through systematic parameter exploration.
• →Machine learning approach resolves long-standing design questions around optimal channel-to-barrier thickness ratios in heterojunction devices.
• →ML-guided TCAD simulation represents hybrid methodology that augments rather than replaces physics-based design validation.
• →Framework applicability extends beyond GaN devices to broader semiconductor design challenges in power electronics and RF applications.