Adaptive Physics Transformer with Fused Global-Local Attention for Subsurface Energy Systems

Key technical wins:

• First neural operator that learns directly from AMR (Adaptive Mesh Refinement) simulations without interpolating to uniform grids
• Gated fusion adaptively balances local (GNO) and global (Perceiver) information at every spatial point
• Latent-dynamics token compression decouples compute cost from input mesh size

Industrial impact: directly deployable for geothermal reservoir extraction and carbon sequestration engineering decisions.