Source-linked AI summary
Deep Learning for Physical Processes: Incorporating Prior Scientific Knowledge
Emmanuel de Bezenac, Arthur Pajot, Patrick Gallinari
TL;DR
The paper asks how Deep Learning can model complex natural processes despite the limitations of purely data-driven approaches. It designs an SST forecasting network guided by a general advection-diffusion solution, and reports performance comparable to a state-of-the-art numerical model while outperforming alternative neural networks.
Problem
Direct Machine Learning methods have not demonstrated competitive performance for complex natural phenomena such as sea surface temperature, despite abundant data.
Method
The paper designs a Deep Neural Network from the solution of a general advection-diffusion PDE, learning a motion field for forecasting SST images.
Results
Performance is comparable to a state-of-the-art numerical model and clearly outperforms alternative neural network models.
Takeaways & Limitations
Physics-derived structure provides a principled design route for Deep Learning models applicable to a class of transport problems following advection-diffusion principles.
Takeaways & Limitations
Motion-field estimation requires optimization and is slower than straightforward neural-network predictions; numerical schemes also require intermediate forecasts to avoid divergence.
Abstract
from arXiv · showhide
We consider the use of Deep Learning methods for modeling complex phenomena like those occurring in natural physical processes. With the large amount of data gathered on these phenomena the data intensive paradigm could begin to challenge more traditional approaches elaborated over the years in fields like maths or physics. However, despite considerable successes in a variety of application domains, the machine learning field is not yet ready to handle the level of complexity required by such problems. Using an example application, namely Sea Surface Temperature Prediction, we show how general background knowledge gained from physics could be used as a guideline for designing efficient Deep Learning models. In order to motivate the approach and to assess its generality we demonstrate a formal link between the solution of a class of differential equations underlying a large family of physical phenomena and the proposed model. Experiments and comparison with series of baselines including a state of the art numerical approach is then provided.
1 Introduction
The paper examines whether physics-informed Deep Learning can model complex natural processes, using sea surface temperature forecasting as a representative task. It proposes a neural architecture guided by advection-diffusion principles and evaluates it against numerical and neural baselines.
- Motivation: Large sensor datasets are challenging the traditional differential-equation framework used to model complex natural phenomena.The paper contrasts physics-based modeling, which encodes conservation laws and principles, with prior-agnostic statistical Machine Learning.
- Application: Sea surface temperature forecasting matters for weather analysis, biological systems, and applications including weather forecasting and coastal planning.
- Research gap: Standard SST forecasting uses coupled ocean-atmosphere systems based on Navier–Stokes equations, while direct ML applications have not produced competitive state-of-the-art results.The physical models rely on multiple hypotheses and do not optimally exploit available data, whereas direct ML methods are also insufficiently competitive.
- Approach: The proposed Deep Neural Network uses a general PDE solution as a guideline for designing a model for SST prediction.The paper introduces a solution class for transport and propagation phenomena, then derives the Deep Learning architecture and compares it with baselines.
- Contributions: The contributions include formal links to an advection-diffusion PDE, unsupervised motion-field estimation, and competitiveness with state-of-the-art numerical methods.
2 Physical Motivation
The paper models SST evolution through an advection-diffusion equation whose solution combines motion and Gaussian diffusion. This solution motivates a neural architecture that estimates motion from data and uses it to forecast future temperature images.
- Physical forecasting: Classical SST forecasting couples PDE-based physical models with satellite data through assimilation to adjust initial conditions before forward integration.
- Advection: The advection equation describes temporal evolution under displacement, and numerical schemes estimate the motion field used for forecasting.
- Advection-diffusion: Advection alone is insufficient for SST, so the transport model combines displacement with diffusion from high- to low-concentration regions.
- Advection-diffusion: The advection-diffusion equation covers physical processes including fluid dynamics, heat conduction, and wind dynamics; D is the diffusion coefficient and ∇2 the Laplacian.
- PDE solution: For suitable initial conditions, the PDE has a unique global solution whose kernel is a Gaussian with mean x − w and variance 2Dt.
- PDE solution: The solution computes future values by convolving the initial condition with a Gaussian when motion w and diffusion coefficient D are known.Because these quantities are unknown, the proposed model learns a motion field analogous to w from data for SST prediction.
3 Model
The model predicts motion fields from past SST images and uses a differentiable, physics-inspired warping scheme to forecast future images. End-to-end training supervises forecasts against target SST images, while autoregressive feedback enables multiple time-step predictions.
- Model architecture: The architecture combines a convolutional-deconvolutional motion estimator with a warping module that forecasts an image from the last input and predicted motion field.The full system is trained end-to-end using only target SST-image supervision, producing an interpretable latent velocity field.
- 3.1 Motion Estimation: Four concatenated SST images are input to the motion-estimation network, whose output is a two-dimensional motion vector for every image pixel.The network uses skip connections, batch normalization, and Leaky ReLU nonlinearities.
- 3.1 Motion Estimation: The predicted motion field is weakly supervised because the warped input image is compared with the next target image rather than with a ground-truth motion field.This discrepancy provides the training signal for the motion estimator through the differentiable warping operation.
- 3.2 Warping Scheme: The warping scheme computes each future pixel as a Gaussian-weighted average of previous-image temperatures centered at the pixel’s estimated previous position.Its kernel is parameterized by the diffusion coefficient D and time step Δt, linking the operation to the advection-diffusion solution.
- 3.2 Warping Scheme: The differentiable warping mechanism is adapted to phenomena governed by advection-diffusion equations and permits error backpropagation into the motion-estimation module.The model is therefore designed to incorporate physical background knowledge while retaining trainability through image-level supervision.
- 3.3 Loss function: The Charbonnier penalty measures forecast discrepancy while reducing outlier influence relative to an ℓ2 norm; a tested Laplacian-pyramid loss reduced generalization performance.The model can also incorporate additional PDE-based knowledge or constraints through penalty terms in the loss.
4 Experiments
The experiments evaluate a physics-inspired SST forecasting model on synthetic ocean data against neural and numerical baselines, using MSE, runtime, and qualitative predictions. The model outperforms the baselines in MSE and is computationally fast, but motion-field estimation adds optimization cost.
- 4.1 Dataset description: The dataset contains 3,734 daily 481 × 781-pixel SST acquisitions generated by the NEMO ocean-modeling framework from 2006-12-28 to 2017-04-05.NEMO uses historical data and data reanalysis to generate synthesized ocean states.
- 4.1 Dataset description: Training and validation use 2006–2015 data, while 2016–2017 test selected 64 × 64 subregions where hot and cold waters interact.The training set contains 94,743 examples, with 20% withheld for validation.
- 4.2 Baseline Comparison: Models are evaluated with mean square error over a six-step forecasting horizon, and neural networks run on a Titan Xp GPU for runtime comparison.Forecasts cover I_t+1 through I_t+6, whose errors are averaged.
- 4.3 Quantitative Results: The proposed model achieves lower MSE than every baseline, including the numerical assimilation model, while direct predictors produce blurrier or less accurate sequences.ConvLSTM captures dynamics better than ACNN, while the GAN baseline reduces blurring but remains below the proposed model.
- 4.3 Quantitative Results: The proposed model is extremely fast, with runtime just above ConvLSTM; the numerical model’s CPU timing is not directly comparable with GPU timings.The numerical assimilation model was run without GPU code, whereas the neural models used a Titan Xp GPU.
- 4.3 Quantitative Results: Estimating the motion field requires an optimization procedure that is slower than straightforward neural-network prediction, and numerical forecasting requires multiple intermediate forecasts to avoid divergence.These computational requirements constrain the otherwise fast approach.
- 4.3 Quantitative Results: Qualitatively, predictions remain close to targets at t + 1 and t + 3 but begin diverging at t + 6, while the model appears to conserve temperatures.Figure 5 compares target, predictions, flow, numerical assimilation, ACNN, and ConvLSTM over daily temperatures from January 17 to January 23, 2017.
5 Related Work
Related work connects the paper to spatio-temporal statistics, optical-flow estimation, and video prediction while distinguishing SST dynamics modeling from object-motion prediction. The paper emphasizes motion-field modeling and physical priors as differences from direct frame-prediction approaches.
- Spatio-temporal statistics: Spatio-temporal statistics uses physical background knowledge and partial differential equations to inspire autoregressive models in hierarchical Bayesian frameworks.This line of work primarily addresses statistical modeling rather than the paper’s deep architecture for SST prediction.
- Video prediction and motion estimation: The paper shares motion-estimation and image-prediction components with computer vision, but its goal is modeling whole SST dynamics rather than tracking moving objects.The distinction concerns both the application domain and the modeled phenomenon.
- Optical Flow: Classical optical-flow methods retrieve apparent motion between consecutive frames using brightness constancy and additional constraints.The extracted motion can support tasks such as object detection, tracking, navigation, and visual odometry.
- Optical Flow: Deep optical-flow methods estimate motion with supervised CNN regression, CNN ensembles, or unsupervised CNN and Spatial Transformer Network architectures.Limited annotated flow data motivates synthetic pretraining or unsupervised approaches.
- Video prediction: Direct next-frame predictors can produce blurred images; GAN regularization sharpens predictions, while Video Pixel Networks improve accuracy at the cost of substantial model and dataset complexity.The paper did not test Video Pixel Network because its code was unavailable.
6 Conclusion
The conclusion presents physics-guided deep learning as a principled alternative for complex natural-process modeling. For SST, the approach generalizes to advection-diffusion problems and compares favorably with neural and numerical baselines.
- 6 Conclusion: The proposed method designs a Deep Learning model for ocean dynamics using inspiration from physics and can generalize to problems following advection-diffusion principles.The paper frames this as cross-fertilization between data-intensive and classical physical modeling paradigms.
- 6 Conclusion: The model reaches performance comparable to a state-of-the-art numerical model and clearly outperforms the alternative neural-network baselines.This conclusion summarizes the reported SST forecasting comparisons.
A Proof of the theorem in section 2 1
The proof applies the Fourier transform to the advection-diffusion equation, solves the resulting first-order ordinary differential equation, and transforms back to obtain a spatial-domain convolution solution.
- The 2D Fourier transformation is applied to the differential equation, using linearity to decompose the transformed left-hand side into three terms.
- The transformed equation becomes a first-order ordinary differential equation with solution f(t) = f(0)e−at.
- The inverse Fourier transform returns the solution to the spatial domain, where frequency-domain multiplication becomes spatial-domain convolution.
- The advection term corresponds to a spatial shift, while the diffusion term is represented by a Gaussian whose variance is identified as σ2 = 2Dt.
B Additional samples from our model
The additional samples show outputs from the model for two four-day periods in May and January 2016, with target, prediction, and flow displayed from top to bottom.
- For 6–9 May 2016, the figure presents the target, model prediction, and model flow from top to bottom.
- For 6–9 January 2016, the figure presents the target, model prediction, and model flow from top to bottom.