Research Areas

PhiLM Lab works toward a more physical AI for a more open world: letting AI understand the physical world beyond shape. Our theme is physical-field reconstruction and learning × world models, from physical-field modeling to autonomous decision making. We combine efficient and explainable AI, world models of radio environments (our current focus), and learning-based optimization (L2O) for real-world communication and network decisions. Our goal is interpretable, editable, and deployable methods that connect physics, data, and decision making.

Three Directions

01 Modeling & Generation — 3D scene representation; physical-field reconstruction and generation, currently centered on wireless RF fields.
02 Efficient Computing & Deployment — AI infra, CUDA acceleration and efficient computation; model deployment and inference optimization.
03 Tasks & Decisions — learning to optimize, reinforcement learning; planning and execution-policy solving.

01 Modeling & Generation 3DGSNeRFRF

3D scene representation and physical-field reconstruction and generation. Our current focus is the wireless radio-frequency (RF) field: physically grounded, data-driven models of radio environments that fuse geometry and materials (e.g., BIM, point clouds, maps), contextual signals (vision, language, IMU, GNSS), and RF measurements (sweeps, arrays). We combine neural representations (3DGS and NeRF variants, diffusion and energy models, GNNs, neural operators such as FNO, DeepONet, and GNO) with physical consistency, constitutive relations, and structured sparsity or low-rank priors. The models support editable inversion, uncertainty quantification, and active re-measurement for localized statistical channel modeling and environment-aware communication.

RF radiance fields and Gaussian splatting for radio environments; structured Bayesian inference on graphs
PDE and boundary consistency; material priors; calibrated uncertainty with multimodal fusion
Geometry / materials plus RF arrays plus contextual signals

02 Efficient Computing & Deployment CUDACompressionFPGA

Structured sparsity, low-rank, and tensorization across data, features, models, and gradients, with unified compression and distillation for edge–cloud collaboration and federated training. AI infra and HW/SW co-acceleration (CPU, GPU/CUDA, FPGA) enable efficient inference and reliable on-device deployment of field models and decision policies.

Structured sparsity and low-rank modeling
Compression and distillation for models and gradients
Communication-efficient training and adaptation
CUDA acceleration, model deployment, and inference optimization
Algorithm–system co-design for throughput and energy gains

Explainable AI Theory

Principled interpretability and reliability for the models we deploy: from attribution and concept-based explanations to causal and statistical guarantees.

Attribution and concept-based explanations; counterfactual and causal probes; manifold learning
Generalization and convergence bounds for nonconvex / structured models
Risk control with conformal prediction; safety constraints and robust training

03 Tasks & Decisions L2OGNNRL

Learning-based optimization (L2O) and reinforcement learning for planning and execution-policy solving. Differentiable, learning-augmented solvers (RL, GNN, and neural combinatorial optimization) operate on the learned world model for constrained decisions, with feasibility guarantees, learned warm-starts, and fast cross-scenario adaptation.

UAV sensing plans, site placement; beam, power, and spectrum allocation; routing and formation planning
Reinforcement learning for planning and execution-policy solving on the world model
Relax–optimize–and–sample; neural projections; constraint satisfaction in deployment
Edge and federated execution for real-time, certifiable decisions

From Representation to Decisions Systems

The three directions form one pipeline: world models + sparse/efficient AI + L2O, coupled with algorithm–hardware co-design (C++, CUDA, FPGA; edge and federated) for online decision making. It maps structure across data, algorithms, and architectures into solvers and hardware.