Dissipative Relaxation Transfer Learning Achieves Data-Efficient Electromagnetic Simulation Accuracy

Researchers are tackling the persistent challenge of accurately modelling resonant phenomena within electromagnetic simulations using neural networks. Sunghyun Nam from the Korea Advanced Institute of Science and Technology, Chan Y Park from KC Machine Learning Lab at KAIST, and Min Seok Jang, alongside their colleagues, present a novel data-efficient training framework called dissipative relaxation transfer learning (DIRTL). This work is significant because it addresses the instability and performance degradation caused by high-amplitude resonances, which often manifest as outlier samples during training. By initially pre-training models with a small amount of fictitious material loss to smooth the response landscape, and then fine-tuning on lossless data, DIRTL enables stable adaptation and substantially improves prediction accuracy , achieving up to a two-fold error reduction in some cases , establishing a physically grounded and versatile approach to enhance machine-learning-based electromagnetic solvers.

Antenna phenomena remains a central challenge. High-amplitude resonances generate strongly localised field patterns that appear as outlier samples, deviating significantly from the general distribution of non-resonant cases, leading to instability and degraded predictive performance. To address this, we introduce dissipative relaxation Transfer learning (DIRTL), a data-efficient training framework that integrates transfer learning with loss-regularised optimisation principles from high-Q photonics. DIRTL first pretrains the model on data generated with a small fictitious material loss, which broadens sharp resonant modes and suppresses extreme field amplitudes. This smoothing of the response landscape encourages stable learning and improved generalisation.

DIRTL training improves resonance prediction accuracy significantly

Scientists have demonstrated that0.1(a)? The grating consists of a 1 × 64 array in free space, illuminated by a transverse magnetic (TM) plane wave normally incident to the device? Each element alternates between free space and a dielectric material with Refractive index nn =? The training and testing datasets span wavelengths from 650nm to 750nm in 10nm increments, with all EM field data generated using the meent rigorous coupled-wave analysis (RCWA) solver? Simulations employed Fourier orders, which were shown to be sufficient for numerical convergence (see Section 1 of the Supplementary Information)?

From this problem setup, we observe the emergence of the resonant outlier cases as described above? As illustrated in Fig0.1(b), most of the generated test samples are non-resonant, not showing pronounced localized field enhancements in the grating space, whilst cases of highly localized fields with large field amplitudes occasionally emerge? This distribution is quantified in Fig0.1(c), displaying a long tail corresponding to strongly resonant cases with maximum amplitudes reaching0.1(c)?. The core challenge addressed is accurately modeling resonant phenomena, which generate outlier samples that degrade predictive performance due to their strongly localized field patterns.

Researchers tackled this by first pretraining models on data generated with a small fictitious material loss, effectively broadening resonant modes and suppressing extreme field amplitudes. This smoothing of the response landscape allows the model to learn global modal features more effectively before fine-tuning on the target lossless dataset containing true high-amplitude resonances. Experiments revealed that applying DIRTL to both Fourier Neural Operator (FNO) and UNet architectures yields substantial improvements in prediction accuracy. Specifically, the FNO variant achieved up to a two-fold error reduction when utilising the DIRTL framework.

The team measured prediction errors across a one-dimensional multi-wavelength binary grating, illuminated by a transverse magnetic plane wave, and observed consistent performance gains with DIRTL compared to standard training methods. Simulations employed 81 Fourier orders to ensure numerical convergence, and the training and testing datasets spanned wavelengths from 650nm to 750nm in 10nm increments. Data shows that DIRTL exhibits robustness across diverse training conditions and supports strong multi-task performance, highlighting its generalizability. The research focused on a 1 × 64 array in free space, alternating between free space and a dielectric material with a refractive index of 2.

Scientists recorded that most generated test samples were non-resonant, lacking pronounced localized field enhancements, while resonant cases exhibited significantly different field patterns. This framework achieves approximately nine-fold sample efficiency, demonstrating its potential as an efficient method for training electromagnetic surrogate solvers. Measurements confirm that DIRTL implicitly embeds physical knowledge into the learning process, offering applicability beyond electromagnetic simulations. The breakthrough delivers a physically grounded and architecture-agnostic curriculum for enhancing the reliability of machine-learning-based electromagnetic surrogate solvers. Tests prove that the two-stage learning process, inspired by photonic optimization strategies, guides the network through a refined learning landscape, enabling smooth convergence and enhanced generalization. Altogether, these results establish DIRTL as a valuable tool for accurately capturing resonant behaviours crucial in photonic device design.

DIRTL improves resonant electromagnetic simulation accuracy significantly

Scientists have developed a new training framework called dissipative relaxation transfer learning (DIRTL) to improve the accuracy of neural network surrogate solvers used in electromagnetic simulations. The core challenge addressed is the difficulty these solvers have with accurately modeling high-amplitude resonant phenomena, which often appear as outlier data points. DIRTL employs a two-stage process, initially training the model with data incorporating a small amount of fictitious material loss to broaden resonant modes and reduce extreme field amplitudes. Subsequently, the pretrained model is fine-tuned on a lossless dataset, enabling stable adaptation and improved prediction accuracy.

Results demonstrate substantial improvements across both Fourier Neural Operator (FNO) and UNet architectures, with the FNO variant achieving up to a two-fold reduction in error. Furthermore, the framework exhibits robustness across varying training conditions and supports effective multi-task performance, indicating its generalizability. The authors acknowledge that the method’s effectiveness relies on the specific curriculum and may require careful selection of the initial loss parameter. DIRTL establishes a physically grounded and architecture-agnostic approach to enhance the reliability of machine-learning-based electromagnetic surrogate solvers. The method’s data efficiency, demonstrated by achieving comparable accuracy with significantly less training data, is particularly noteworthy. Future research could explore the application of DIRTL to other physical systems where physics-informed learning is crucial, potentially broadening its impact beyond electromagnetic simulations.