RC thermal simulator too smooth for GNN to outperform LSTM, how to design a simulation where spatial graph structure genuinely matters?
Building a GNN vs LSTM comparison for thermal prediction in an immersion-cooled server rack. Using a lumped RC model:
C_i * dT_i/dt = Q_i(u_i) - (T_i - T_fluid)/R_conv + sum_j[(T_j - T_i)/R_ij]
After 300 samples and 80 epochs, GNN, LSTM, and GNN_NoEdges (ablation with empty edge index) all converge to within 0.03°C MAE of each other. Removing all graph edges makes essentially zero difference.
My hypothesis: the RC ODE is dominated by the local term. Each server's next temperature is ~92% determined by its own previous temperature and load. The neighbour coupling term is too weak relative to self-dynamics for message passing to add anything beyond what a per-node LSTM already learns.
Specific questions:
- Is this diagnosis correct, is the RC model's linear self-dominance the root cause?
- What simulator design choices would make spatial propagation the dominant factor rather than self-dynamics? Specifically: what R_neighbor / R_conv ratio would make neighbour coupling matter enough for a GNN to win?
- Is there a class of thermal problems where GNNs demonstrably outperform LSTMs in the literature? (chip thermal maps, CFD surrogate models, heat exchangers?)
- Would switching to a nonlinear thermal model (e.g. radiation terms, phase-change immersion cooling) create enough spatial complexity for graph structure to matter?
Rack config: 16 servers, linear topology, TDP 350-720W per server (non-uniform), asymmetric convective resistance, hotspot injection at 8% probability per step.