- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
This work proposes Physics-Defined Energy-Matching Diffusion (PhysDEM) as a way of retrieving the complete map of coefficients and their associated solution fields from sparse and noisy measurements. The method was tested on several PDE problems and applications involving reservoir permeability and pressure reconstruction, as well as a contamination transport problem using the Bemidji dataset. The main contribution is a theoretical construction that proposes pairs of coefficients and their associated solutions through a linearized model and constructs a reference for the coefficients-solution probability distribution using a Gaussian energy function and learns, through a diffusion model, the correction residual between the reference and the complete distribution, using data generated by MCMC chains rather than a premade dataset, resulting in a model that can produce alternative reconstructions without the need for retraining.
Critical strengths
- I believe that the method and the proposal are an elegant solution to model coefficient and solution fields jointly. The formulation provides a way to avoid solving complex systems that may be easy to evaluate but hard to solve with standard methods, offering an alternative in such cases.
- The design choices for the method are theoretically well justified. The authors propose the reference distribution analytically and derive the necessary correction to be learned. They establish how the sampler should use the correction and the learning objective, connecting the original problem of learning the complete distribution to residual learning from this reference, giving the neural network a defined role rather than treating the model as a black box.
- The trained sampler can be used across multiple configurations without retraining.
- The paper is well written and easy to understand.
Critical weaknesses
- Despite the elegant formulation, I find the evaluation of this work flawed as it is mainly statistical and not consistent across all experiments. Furthermore, apart from the results in the appendix, the evaluation does not report reconstruction errors of the sampled coefficients and their respective solutions. The results in the appendix are also concerning; relative errors of 1.028 and 1.019 can be obtained with a near-zero prediction, showing that the method, in my opinion, fails in these problems. I find the authors' justification of weak constraints and that further examination should be done regarding this.
- Related to the previous point, I believe that the statistical metrics are not enough to support the claims about the generated fields. These metrics have substantial flaws. For instance, spatial summaries can hide misplacement issues and average rank can balance if you have the same amount of overprediction and underprediction. These can be useful to characterize the learned distribution but should not replace a full field recovery metric set.
- The benefit of the correction is unclear over the reference. The paper shows that the spread-skill improves substantially from the reference distribution by incorporating the correction. However, it does not specify the component responsible for this improvement as the skill can improve by either improving the reconstruction error because the spread increases or both.
- The coarseness of the grid used is also concerning. All the experiments use a spatial grid with 48 time points, without a grid study. Training depends on generating good initial solutions using pCN across multiple configurations. The appendix describes all configuration parameters but does not examine how many chains are necessary to train the models on a finer grid, as the model's applicability is currently limited in its current settings.
- The method cites as one of its advantages the lack of pre-computed solutions to train the model. This is appealing, as I said before; however, they fail to quantify the difference between computational cost and performance degradation. Avoiding forward solves is potentially useful only if it produces more accurate solutions, or the cost of solving the system is high enough that the time saved justifies the error. None of this is quantified in the paper.
Questions for authors
- Could you provide a systematic comparison between the proposed method and the baselines on the full field coefficients and solution reconstructions against a FEM reference? Use normalized error.
- Could you compare the reconstruction errors before and after the proposed correction?
- Could you provide for one of the systems in the paper an ablation on grid size, reporting how the training cost and accuracy change?
- Could you report how many pCN chains you require to achieve your results and also how much it increases as the grid becomes finer?