DyPSI: Dynamic Physics Sensing via Joint Field and Sensor-Trajectory Generation

AI summary

Overview

Research area: Generative AI for physical sensing — specifically, joint generative modeling of spatiotemporal physical fields and mobile sensor trajectories (a problem at the intersection of diffusion models, functional tensor decompositions, and optimal experimental design).

Technical level: Advanced. The paper assumes familiarity with diffusion models (EDM formulation), Tucker tensor decompositions, Gaussian processes, and A-optimality from experimental design.

Scope: DyPSI is a conditional generative framework that, given sparse historical sensor readings of an evolving physical field, simultaneously generates the future field and the future positions of the mobile sensors that will observe it.

What This Paper Is About

Reconstructing a dense, time-evolving physical field (a turbulent flow, an ocean temperature map, a pressure field over a car body) from only a handful of point sensors is an ill-posed inverse problem, and it becomes harder when the sensors are mobile and must decide where to move next. Existing learning-based methods handle placement either with a single fixed layout shared across all frames — which cannot follow structures that drift away — or by re-optimizing sensor positions independently at every frame, which produces temporally discontinuous "teleporting" jumps that real platforms cannot execute. DyPSI reframes dynamic sensing as one joint conditional generation problem: predict the future field and the future sensor trajectory together, as a single coherent sample conditioned on the observation history.

Key Contributions

1. Problem reformulation. Dynamic active sensing is recast as joint conditional generation over future fields and future sensor trajectories, p(P_{t+1:T}, s_{t+1:T} | H_t), replacing both static-layout and per-frame placement abstractions with sequence-level modeling.

2. A shared latent representation for two incompatible data types. Discrete sensor coordinates are lifted into continuous Sensor Position Fields (sums of Gaussians), and both fields and sensing configurations are encoded with a Functional Tucker decomposition whose spatial bases are shared across modalities, then concatenated along the channel mode into one joint latent (R_joint = R_1^phy + R_1^sen).

3. A sequence-aware joint diffusion model. A single denoiser is trained on that joint latent, with the standard i.i.d. EDM noise replaced by a Matérn-3/2 Gaussian-process-correlated perturbation along the time mode — adding temporal smoothness pressure without adding a regularizer term to the loss.

4. Two theoretical consistency checks plus three-benchmark validation. Proposition 1 bounds the sequence-aggregated A-optimality criterion (sum of posterior covariance traces) by the idealized EDM denoising risk up to constants C_low, C_high; Proposition 2 bounds the discrete trajectory variation sum ||s_{t+1} - s_t||^2 by 2 c_K L_GP(s), with c_K = λ_max(D K D^T).

Main Findings

• Lowest error on every benchmark and sensor budget. Against the strongest static baseline (PhySense-opt), DyPSI reduces relative L2 error by 10.87% to 15.35% on turbulent flow (#200 to #30 sensors), 8.43% to 9.84% on sea surface temperature (#100 to #15 sensors), and 17.47% to 19.43% on car aerodynamics (#200 to #15 sensors). The authors report that gains generally become more pronounced as sensor availability decreases.

• Fewer sensors beat more sensors. On the car mesh, 15 dynamically coordinated sensors reach an error of 0.0311, outperforming PhySense-opt with 30 sensors and PhySense with 100 randomly placed sensors.

• Statistical significance reported. All turbulent flow and car aerodynamics improvements are significant at p < 0.001; sea surface temperature improvements at p ≤ 0.01 (five independent random seeds; full standard deviations and paired t-test values in Appendix D).

• Kinematic feasibility is the distinctive win. Compared against two oracle strategies that see future ground truth and run test-time optimization:

  • Per-frame Oracle achieves the lowest error (0.1007 turbulent flow with 50 sensors; 0.0259 sea temperature with 25 sensors) but violates the velocity limit in 89% and 75% of transitions, with smoothness costs of 142.5 and 165.3.
  • Sequence Teacher reaches 0.1206 and 0.0355 with only 5% and 2% violation rates, but costs 1683 and 1429 seconds per 64-frame sequence.
  • DyPSI reaches 0.1352 and 0.0394 — within 0.0146 and 0.0039 of the Sequence Teacher — with violation rates of 8% and 4%, smoothness costs of 4.2 and 9.2, and runtime of 1.9 and 2.1 seconds.

• Joint generation beats two separate models. On turbulent flow with 50 sensors, a joint model lowers relative L2 error from 0.1421 ± 0.0045 to 0.1352 ± 0.0024 and smoothness cost from 9.1 ± 1.1 to 4.2 ± 1.0, compared with separate diffusion models sharing only the history encoder.

• Matérn-3/2 is the right temporal kernel. i.i.d. noise gives 0.1515 ± 0.0039 error and 18.6 ± 1.9 smoothness; RBF gives 0.1489 ± 0.0025 and 6.2 ± 1.3; Matérn-3/2 gives 0.1352 ± 0.0024 and 4.2 ± 1.0 — RBF's smoother prior over-restricts necessary temporal variation.

• The learned strategy transfers across reconstruction backbones. Using Senseiver, PhySense, or DiffusionPDE independently as the offline teacher, DyPSI-opt achieves the lowest error at every budget. With 30 sensors, error drops from 0.4570 to 0.4527 (Senseiver), 0.2157 to 0.1826 (PhySense), and 0.7904 to 0.7116 (DiffusionPDE). Grid and Min-Max sampling do not consistently beat random placement, showing spatial regularity alone is insufficient.

• Qualitative behavior matches the intent. Three generated trajectories follow coherent flow structures over frames 0–9 without abrupt jumps; static optimization retains errors near structures that have advected away from fixed sensors, and DyPSI repositions sensors around high-gradient filaments. Sea temperature trajectories respect the land mask.

Methodology in Plain English

The core mismatch. A physical field is a continuous function over space; a set of sensor positions is a short list of coordinates. You cannot feed both into one network without translating them into a common language.

Step 1 — Turn coordinates into a picture. Each sensor layout is converted into a smooth heat map called a Sensor Position Field: a sum of Gaussian bumps centered at the sensor locations (with widths σ_x, σ_y, extended as a Gaussian product in 3D). This is treated as a differentiable spatial surrogate, not a new physical variable. Going back from heat map to coordinates at inference is a top-k argmax readout.

Step 2 — Compress both modalities with one shared basis. Time is split into segments of length T_seg. Both the field sequence and the Sensor Position Field sequence are compressed with a Functional Tucker decomposition — a tensor core multiplied by learned coordinate basis functions, where the basis functions are coordinate MLPs. The spatial bases are trained once on the whole dataset and shared across both modalities and all samples, justified because both objects live on the same domain Ω; only the channel-mode rank differs. The resulting core tensors are concatenated, so a single denoiser attends to field and sensors at every layer.

Step 3 — Diffuse jointly, with temporally smooth noise. Standard diffusion training corrupts data with independent random noise at each frame, so the denoiser learns to fix each frame in isolation and has no reason to produce smooth motion. DyPSI instead draws one noise sample correlated across time using a Matérn-3/2 kernel, implemented via its Cholesky factor. The loss form is unchanged EDM, but the denoiser must now recover the clean latent from a temporally smooth corruption, which biases both the generated field and the generated trajectory toward temporal coherence. No extra regularizer is added.

Step 4 — Sample conditioned on history. At inference, the observed prefix frames are softly anchored during reverse denoising, blending the current noisy estimate toward the perturbed historical latent with a weight α_r that anneals from 1 to 0. After sampling, both modalities are decoded by inverse Functional Tucker and sensor coordinates are read out from the SPF channel.

Training trajectories come from a costly offline sequence-level optimizer with access to ground-truth fields; PhySense is the default teacher reconstructor. At test time DyPSI uses no ground truth and no test-time optimization.

Why This Matters

Impact on research. The paper supplies a concrete recipe for the representation problem in physics-informed generative modeling — how to put heterogeneous objects (dense fields, sparse coordinate sets) into one latent without architectural duplication — and it makes the case that placement should be modeled as a sequence rather than a per-frame optimization. The two propositions give the design a stated (if idealized) theoretical footing: the reconstruction-based training signal is tied to cumulative posterior uncertainty, and the temporal kernel is tied to bounded trajectory variation. The runtime comparison (roughly 1.9–2.1 seconds versus 628–1683 seconds for optimization-based alternatives) reframes adaptive sensing as an amortized inference problem rather than an online optimization one.

Real-world applications named in the paper:

• Ocean and climate monitoring — drifting buoys and autonomous underwater vehicles tracking sea-surface temperature; the GLORYS12 experiments explicitly mask continents so trajectories stay in valid ocean regions.
• Environmental fluid dynamics — following drifting vortices, propagating fronts, and turbulent structures with mobile sensing platforms.
• Aerial and autonomous platforms — repositioning airborne sensors over time to track evolving phenomena.
• Vehicle and industrial design — surface-pressure reconstruction over car bodies, where 15 well-placed moving sensors outperform 100 randomly placed ones; the paper also cites industrial inspection as a target domain.

Industry relevance. The value proposition is sensor economy: the car-aerodynamics result shows dynamic coordination cutting required sensor count by roughly an order of magnitude at matched accuracy. Combined with sub-second-per-sequence inference and explicit velocity/acceleration feasibility handling, this is closer to deployable than optimization-based planners, which need minutes to hours per sequence.

Future Directions

• Closing the loop. The current evaluation is a history-conditioned open-loop rollout, not fully closed-loop sensing; the authors note that low inference cost permits receding-horizon re-planning, which remains untested.
• Removing teacher dependence. Training trajectories come from an offline optimizer with ground-truth access, and DyPSI does not fully match that oracle teacher. Methods that reduce or eliminate this supervision gap are open.
• Richer observation and dynamics models. The formulation assumes point-wise observations, whereas real sensors such as radar integrate over small neighborhoods; and feasibility covers only velocity/acceleration limits, land masking, and surface projection — not underactuation or ocean currents. Both extensions are explicitly called out.
• Robustness and scale. Behavior under abrupt regime shifts, strongly non-Gaussian sensing noise, and sensor densities far above those tested (where SPF peaks may merge and the top-k readout needs the non-maximum-suppression fallback) is unaddressed, and no out-of-distribution claims are made.

Target Audience

Researchers and graduate students working on generative models for spatiotemporal physical data, physics-informed machine learning, and data-driven sensor placement or optimal experimental design. It will be most useful to readers already comfortable with diffusion models and tensor decompositions who want a template for jointly generating heterogeneous modalities with shared structure. Practitioners in oceanography, aerodynamics, and autonomous sensing will find the benchmark results and runtime numbers relevant, but the theory sections require the advanced background noted above.

Authors’ abstract

Physics sensing in evolving environments requires reconstructing dense spatiotemporal fields from sparse observations while adapting where future observations should be collected. Existing reconstruction and sensor-placement methods have made substantial progress, but they typically rely on static sensing assumptions: they either use a single layout across all frames or optimize placements independently at each frame, which limits their ability to produce temporally coherent sensing trajectories. We introduce DyPSI, a dynamic physics sensing framework that formulates this problem as the joint generation of future fields and sensor trajectories conditioned on historical observations. DyPSI lifts discrete sensor coordinates into continuous Sensor Position Fields, encodes fields and sensing configurations with a shared Functional Tucker representation, and trains a joint diffusion model to generate future fields and sensing trajectories simultaneously in the resulting latent space. A Gaussian-process temporal kernel correlates the perturbation injected during diffusion training, biasing the denoiser toward temporally smooth recoveries, and we further provide an analysis linking the EDM training objective to a sequence-aggregated A-optimality criterion and a kernel-induced bound on discrete trajectory variation. Experiments on turbulent flow, GLORYS12 sea-surface temperature, and 3D car aerodynamics show that DyPSI consistently outperforms static and frame-wise placement strategies, with substantial error reduction under tight sensor budgets. The code is available at https://github.com/mengqi132/DyPSI.