acceptodds
Under review as a conference paper at ICLR 2027

PHYSINET:PHYSICS-GUIDED WEAKLY-SUPERVISED DEEP LEARNING FOR INTERPRETABLE SEISMIC EVENT CLASSIFICATION AND INVERSE PHYSICAL DISCOVERY

Abstract

Deep seismic classifiers are fast but opaque and fragile under distribution shift, while classical physical discriminants are interpretable yet require manual mea- surement at deployment. We propose PhysiNet, a physics-guided weakly- supervised framework built on one principle: physical knowledge constrains training, while inference uses only three-component time–frequency images and no auxiliary physical quantity. We instantiate the framework on the three-class earthquake–explosion–collapse task with two auxiliary targets chosen for their source physics rather than their availability: the classical lg(P/S) amplitude ra- tio — converted, to our knowledge for the first time, into a weak regression tar- get — and the P–S spectral centroid. Both quantify the radiated wavefield (a dimensionless amplitude ratio and a normalized frequency statistic) rather than propagation geometry, and are delivered as masked regression targets over a shared encoder. On DiTing 2.0, PhysiNet improves three-class event-level ac- curacy from 0.9017 to 0.9178 over a matched classification-only baseline (means over ten training seeds), each target contributing about one point alone and +1.6 points jointly. Strikingly, the lg(P/S) head attains only moderate regression fidelity (R2 =0.521): the value of weak supervision lies in constraining the representa- tion, not in measurement precision. Beyond accuracy, the trained model becomes an instrument for inverse physical discovery: aggregating its auxiliary outputs over the test population reveals class-conditional radiation regimes in the lg(P/S)– centroid plane, a physically grounded ambiguity map that co-locates residual er- rors with class boundaries, and a previously unquantified within-class trade-off between P/S energy partition and characteristic radiation frequency — within each source class, records that channel relatively more energy into P also radiate lower- frequency energy. These findings illustrate a bidirectional loop in which physics guides learning and the calibrated model, in turn, quantifies source physics at cat- alog scale.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
seismic event classification, physics-guided deep learning, weak supervision, interpretability, P/S amplitude ratio, inverse physical discovery

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_N1FcOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes PhysiNet, a weakly supervised framework for three-class seismic event classification (earthquake, explosion, and collapse). The method uses a shared ResNet18 encoder with one classification head and two auxiliary regression heads predicting lg(P/S) and P–S spectral centroid. The total objective combines the standard classification cross-entropy loss with MSE losses on the available auxiliary targets.

Critical strengths
  1. The paper addresses a practically relevant seismic classification problem and evaluates earthquake, explosion, and collapse jointly. 2.3 The experiments include useful matched ablations: classification-only, lg(P/S)-only, centroid-only, and the joint model. Both auxiliary targets provide measurable improvements over the baseline.
  2. The auxiliary quantities are physically interpretable.
Critical weaknesses
  1. The methodological contribution appears limited and is essentially conventional multi-task learning. My main concern is that the proposed framework is technically simple. The architecture similarly consists of a shared ResNet18 representation followed by classification and regression heads. Therefore, the central mechanism is very close to standard multi-task learning with additional regression supervision. There is no explicit physical constraint, governing equation or physics-based regularization in the optimization objective. The “physics guidance” mainly arises from choosing physically meaningful quantities as auxiliary labels.
  2. The “inverse physical discovery” claim appears stronger than what the experiments establish. The paper emphasizes that the auxiliary heads can be used for inverse physical discovery. However, Section 4.5 explicitly states that the reported discoveries are first established using ground-truth bulletin targets computed directly from picks and waveforms, and the model-estimated plane is subsequently compared against this ground-truth structure. The physical patterns are already observable by directly analyzing the constructed lg(P/S) and centroid targets without PhysiNet.
  3. The “weakly supervised” characterization is questionable. The primary classification task is fully supervised, and the spectral-centroid auxiliary target is available for all retained samples. The method is essentially standard multi-task learning with partially missing auxiliary labels rather than a distinct weak-supervision methodology.
Questions for authors

Could the authors clarify what specifically makes the proposed framework “weakly supervised”? The primary classification task is fully supervised, and the spectral-centroid target is available for all retained records. How does this setting differ conceptually or methodologically from standard multi-task learning with partially missing auxiliary labels?

Reviewer_h9DeOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces a new approach to classifying seismic wave events, PhysiNET, which uses physical quantities (the logarithmic P/S amplitude ratio and the P-S spectral centroid) during training and then replaces the trained decoder head with a classifier to ensure physical consistency and to mitigate distribution shift. The approach is then evaluated on the DiTing 2.0 dataset comparing PhysiNet with a classification-only baseline and two variants using each physical quantity separately. The evaluation also includes an outline of the comparison against prior classifiers on the DiTng 2.0 three-class task although this is limited to a largely qualitative description of the differences in these approaches in Table 5.

I would like to qualify my review below noting that this paper is slighly outside my area of expertise and I am basing the review on my broad expertise in physics-base ML.

Critical strengths

It appears that the DiTing 2.0 three-class task is a standard benchmark which is accepted in the community and thus provides a worthwhile problem to solve. I think the presentation of the paper is reasonably clear and certainly the motivation of the contribution is nicely conveyed. I believe their is genuine value in incorporating physical constraints and knowledge into physics classifiers to mitigate distributional shift and it appears that the paper is able to do this fairly successfully. The motivation of the physical quantities used is also well-presented and the paper on the whole reads very coherently.

Critical weaknesses
  1. I think my main concern is that I find it difficult to ascertain the impact, novelty and importance of this contribution. If I was to narrow down the paper to its central contribution I would argue it is merely a slightly different composition of existing architectural components (see beginning of section 3) and a novel loss function which includes physics-based terms. It is not clear to me how this might generalise beyond the very specific classification task studied in this work. In particular, it might be helpful if the authors were able to demonstrate that this methodology can be applied to other domains, for example if there are other seismic classification tasks beyond the one considered here and what type of physical quantities would have to be used in this case.

  2. The second main concern I have is that, if we accept the task (DiTing 2.0 three-class) as a standardised benchmark it is rather suprising to see that only two other accuracy numbers are reported in Table 5 beyond the PhysiNet. In my opinion this warrants significant further clarification because at the moment it seems that either the authors are unable to report comparative scores on a widely accepted problem set because they have not investigated the literature thorougly, which is perhaps even more worrying, the problem studied in this particular paper is indeed very narrow, so much so that there are no other comparative baselines that could be reported. Either way I think this might very seriously impact whether this contributions merits publication at ICLR.

  3. A diagramatic representation of PhysiNet would be very helpful to fully understand the replacement of the classifier head.

  4. On the bottom of page 8 there is a missing section reference and on the penultimate paragraph of page 3 the reference to an Appendix does not render correctly.

Questions for authors
  1. Please see point 2 above.

  2. I wonder if it might be possible to measure robustness to distributional shifts in some way? I assume this is difficult based on a single dataset but it would certainly help convince readers of the usefulness of PhysiNet.

Reviewer_12X5OpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces PhysiNet which is physics-guided seismic event classifier. It has two auxiliary regression targets, the lg(P/S) amplitude ratio and the P-S spectral centroid, so that physical knowledge shapes the representation during training. During inference, it needs only three-component time-frequency images and no measured physical quantity.

Critical strengths
  1. The model constraints the representation at the time of training and does not require any measurement at the time of deployment. I think this is a clean way to use the domain knowledge.

  2. I like the four-group design choice, with the matched classification-only baseline along with each target alone. This supports both the leave-one-out and marginal views of each of the auxiliary task's contribution.

  3. The Figure 1 confusion matrices are quite informative. It shows that roughly 90% of the improvement can lands on the explosion-collapse pairs (which lacks an established classical discriminant). This is more convincing than the aggregate number.

Critical weaknesses
  1. The paper describes training on ten seeds but no variability numbers are reported for any of the classification results. Also, in figure 2, the caption reports lg(P/S) R² = 0.726 but the figure’s own plot title and Table 2 both give 0.521. Understanding the standard deviation atleast on the headline numbers will help improve the understanding of the results significantly

  2. Both the described auxiliary targets may still encode geometry. Both use windows derived from Sg−Pg despite the paper’s aim of avoiding geometry-based supervision. The higher earthquake centroid is also attributed to longer windows in the paper. I could not find the promised geometry-supervised control. These controls are helpful to support the source-physics interpretation.

  3. The inverse-discovery claim seems overstated. Within-class correlations are weak (|r| ≤ 0.093), and the explosion result is not significant (p = 0.18). Controlling for class does not account for distance or attenuation. Also, finding no significant difference between classes does not establish a shared source mechanism. The conclusion should acknowledge these alternative explanations more clearly.

  4. The paper has some grammatical errors, typos and missing references (e.g. “Appendix ??” on page 3 and “Section ??” on page 8)

Questions for authors

See weaknesses

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.