LG AI Research Unveils 'GEAR' and 'Sim-GEAR' to Address Data Scarcity Using Riemannian Geometry

LG AI Research announced GEAR and Sim-GEAR, new Transfer Learning methodologies designed to solve data-scarce scientific problems, at ICML 2026. In the field of AI for Science, critical physical properties, such as those involved in new material development, suffer from extremely limited training data due to the high cost and time required for measurement.

GEAR connects the Latent Spaces of different physical property prediction models from the perspective of Riemannian geometry. Overcoming the limitations of traditional Euclidean distance comparisons, it trains to align the Ricci Scalar, which represents the geometric structure of the two spaces, thereby transferring knowledge from a data-rich Source Model to a Target Model. To achieve this, the research team derived and optimized differential equations directly to address memory issues in PyTorch Autograd.

The follow-up study, Sim-GEAR, simplifies GEAR's complex curvature calculation process through Metric Matching. Unlike the Ricci Scalar, which compresses directional information, Sim-GEAR directly aligns the Metric to preserve distance and angle information. This approach eliminates high-order derivatives, reducing computational complexity while maintaining stable performance.

Experimental results show that Sim-GEAR improves computational speed by approximately 91.1% and reduces memory usage by approximately 62.6% compared to GEAR. This enables the processing of larger Latent Dimensions and Batch Sizes, demonstrating the potential of AI technologies that operate efficiently even in data-limited environments. GEAR was accepted to the main conference, while Sim-GEAR was accepted to the AI for Science workshop.