acceptodds
Under review as a conference paper at ICLR 2027

Predicting Asteroid Physical Properties from Photometric Data Using Machine Learning

Abstract

Accurately determining the physical parameters of asteroids, such as shape, rotation period, and rotation axis orientation, is crucial to understanding their structure and evolution, as well as to planning space missions. Since direct observation of the shapes of these objects is usually impossible, these parameters are determined indirectly through photometric observations. This paper describes a method to support this process using machine learning techniques. The final model developed - a two-branch convolutional neural network combining light-curve analysis with geometric metadata - achieves a coefficient of determination (R²) of about 0.40 on an independent test set, a result significantly above the random baseline.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
photometric inversion, asteroid light curves, deep learning, shape modeling, cellinoid model, convolutional neural networks, inverse problems

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_GUAFOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a method for estimating the physical parameters of asteroids from photometric observations. Specifically, the method takes light curves together with geometric metadata as input and employs a neural network to estimate asteroid shape parameters. The model is trained and evaluated on a synthetically generated dataset.

Critical strengths

The problem studied in this paper is interesting and potentially valuable.

Critical weaknesses

In its current form of the paper, I believe there remains a substantial gap between the methodology, experimental validation, and overall presentation of this work and the standard expected for an ICLR publication. My main concerns are as follows.

  1. The assumptions underlying the observation model may be overly simplified. Although the problem itself is interesting, asteroid photometric observations are likely influenced by factors beyond those considered in the paper, such as asteroid shape, the Sun-target unit vector, and the observer-target unit vector. For example, surface material composition, albedo variations, surface scattering properties, and other physical characteristics may also substantially affect the observed brightness. Ignoring these factors may introduce a considerable gap between the synthetic setting considered in the paper and real-world observations. The authors should more carefully justify the assumptions behind the synthetic data generation process and, ideally, investigate how robust the proposed method is when these assumptions are violated.

  2. The practical availability of the geometric metadata is unclear. In addition to the light curve, the proposed method requires a 9-element geometric metadata vector as input. However, the paper does not sufficiently explain how these quantities would be obtained in a realistic deployment scenario. Are all of these metadata directly available from astronomical observations or ephemeris information, or do some of them themselves need to be estimated? If the latter is the case, the accuracy and availability of these quantities could significantly affect the practical applicability of the proposed method. This point should be clarified, and the sensitivity of the method to errors in these metadata would also be useful to evaluate.

  3. The evaluation metrics are not used consistently in the experimental comparison. Section 5.1 introduces several evaluation metrics, including MSE, MAE, and R2R^2. However, the comparison with baseline methods appears to rely only on R2R^2. It is unclear why the other reported metrics are omitted from the baseline comparison. Providing MSE and MAE for all methods would enable a more comprehensive and fair assessment of their relative performance.

  4. The baseline comparison in Table 3 requires substantially more discussion, and the results do not clearly demonstrate the advantage of the proposed method. The descriptions and implementation details of the baseline methods in Table 3 are insufficient, making it difficult to assess whether the comparison is fair and reproducible. More importantly, the "Random Forest + FFT" baseline appears to offer a particularly strong alternative to the proposed approach: it achieves a higher R2R^2 while also requiring less training time. Although its inference time is higher than that of the proposed method, the reported latency is still only on the order of milliseconds, which may be negligible in many practical applications. Based on the current results, it is therefore unclear what practical or methodological advantage the proposed neural-network-based approach provides over this baseline. The authors should provide a more detailed analysis of this trade-off and more convincingly demonstrate the situations in which the proposed method is preferable.

Questions for authors

Please see the weaknesses.

Reviewer_W6oKOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper investigates the direct prediction of asteroid physical properties from photometric light curves using a dual-branch neural network. A 1D CNN processes a uniformly sampled light curve, while an MLP processes geometric metadata, and the fused representation is used to regress five normalized Cellinoid shape parameters. The authors construct a large synthetic dataset and report approximately R^2=0.40for single-light-curve prediction on an independent set of 11 unseen asteroid shapes, while averaging predictions from multiple viewing geometries increases the per-shape R^2 to approximately 0.85. The problem is scientifically relevant, but I have substantial concerns regarding the methodological novelty, the mismatch between the claimed sparse-photometry setting and the actual experimental protocol, the physical information assumed available at inference time, and whether the current experiments are sufficient to support the reported generalization claims.

Critical strengths
  1. The paper considers a scientifically relevant inverse problem and motivates it in the context of large astronomical surveys. Replacing object-specific iterative inversion with amortized neural inference could be practically useful if it generalizes to realistic observations.
  2. The Cellinoid representation is compact and physically interpretable. Predicting five normalized shape parameters is substantially easier to interpret than predicting a high-dimensional point cloud, and the resulting representation can in principle be converted back into a physical surface.
  3. The discussion of parameter observability is useful. In particular, the paper recognizes that normalization removes information about the polar dimensions, and that these variables are therefore inherently difficult to recover from the current input representation.
Critical weaknesses
  1. Limited methodological novelty. The proposed model consists of two standard 1D convolutional layers for the light curve, a single fully connected branch for geometric metadata, feature concatenation, and two dense regression layers, trained using standard MSE and Adam; the complete model contains approximately 111k parameters. The main distinction from prior neural asteroid inversion methods therefore appears to be the choice of a compact Cellinoid output representation rather than a fundamentally new learning mechanism.

  2. Mismatch between the stated sparse-photometry problem and the evaluated task. The experimental setting does not sufficiently match the sparse and irregular photometry problem used to motivate the paper. This is a central claim-evidence mismatch because the task evaluated in the paper is materially easier than the task described in the motivation, and experiments using realistically sparse and irregular observation patterns are necessary before the claimed applicability to survey photometry can be supported.

  3. Potential use of unavailable physical information at inference time. It is unclear whether the model assumes access to physical quantities that are themselves normally part of the asteroid inversion problem. The paper should therefore state precisely which physical quantities are assumed available before applying the network and explain how these quantities would be obtained for a previously uncharacterized asteroid. Otherwise, the current formulation may be solving only a restricted subproblem of the inversion task presented in the introduction.

  4. The multi-geometry generalization result is statistically fragile and insufficiently isolated. The main empirical claim that multi-geometry aggregation substantially improves prediction accuracy is not sufficiently supported by the current evaluation. The independent test set contains only 11 asteroid shapes, while approximately 95 predictions corresponding to different geometries are averaged for each object, increasing the reported R^2 from roughly 0.40 for individual light curves to approximately 0.85 after aggregation. The independent test set of 11 shapes is too small to support broad statistical claims.

  5. Baselines and reported headline performance do not adequately support the claimed advantage. The empirical comparison is not sufficiently strong to establish the value of the proposed architecture, and the reporting of the headline result is potentially misleading. The paper should consistently report the multi-seed mean and variance in the abstract, main text, and conclusion, and the baseline study should isolate the respective contributions of the light-curve branch, metadata branch, fusion strategy, and temporal representation before the claimed superiority of the proposed architecture can be established.

Questions for authors
  1. Section 4.1 refers to the post-fusion fully connected layers as a “deep classification block,” although the task is regression.
  2. The description of the “9-element” metadata vector should be made explicit. Based on the listed geometric variables and sine/cosine encoding, the dimensionality is not immediately obvious from the text.
  3. The statement that dominance of the second Fourier harmonic “confirms physical validity” is too strong. It verifies one expected signal property but does not establish the fidelity of the simulator to real asteroid photometry.
  4. The paper should distinguish consistently between “random baseline” and “mean baseline.”
  5. The 80:20 standard split needs a precise statement of the splitting unit. If individual light curves from the same asteroid shape can appear in both training and test partitions, this benchmark measures view interpolation rather than shape generalization.
Reviewer_K8ZcOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates whether neural networks can replace object specific iterative photometric inversion with direct inference for large scale asteroid surveys. The method predicts a compact and physically interpretable Cellinoid representation consisting of six semi axes, with one fixed for scale invariance and the remaining five estimated by the network.

The model is trained entirely on synthetic Cellinoid light curves generated under diverse illumination and viewing geometries. The raw dataset contains more than 10 million tabular rows, while each observational series is converted into a 72 point light curve sampled over a full rotation. A dual branch convolutional network jointly processes the light curve and a nine dimensional geometry vector and outputs the five normalized shape parameters.

Evaluation includes an independent test set of 11 unseen shapes with 1,046 light curves and a conventional 80:20 split. A favorable run achieves an R squared of about 0.40 and an MAE of about 0.08 for individual light curves. However, the three seed average reported in the model comparison is lower, with an R squared of 0.290. Averaging predictions across approximately 95 viewing geometries per object substantially improves the per object R squared to 0.835.

Additional experiments show only modest improvement when increasing the number of training samples from 58,731 to 673,607. The results also indicate that the parameters b1 and b2 are more observable from light curve amplitude, whereas c1 and c2 are poorly constrained because normalization removes much of the absolute brightness information associated with the spin axis dimensions. Overall, the study suggests that multi geometry information is critical for reliable shape estimation and that simply increasing the number of isolated light curves provides limited benefit. The independent test set contains only 11 shapes, so the reported generalization results should be interpreted as preliminary evidence.

Critical strengths

S1. Practical problem formulation and efficient inference. The paper adopts a compact, physically interpretable Cellinoid representation instead of dense shape reconstruction, enabling efficient prediction and subsequent generation of asteroid surfaces and synthetic light curves. The reported inference time of 0.12 ms per light curve demonstrates the computational efficiency of the approach and its potential suitability for large-scale asteroid surveys.

S2. Transparent reporting of experimental variability and baseline comparisons. The paper explicitly reports variability across random seeds and acknowledges that the single-curve R² of 0.40 represents a favorable run. Table 3 shows that the CNN achieves an average R² of 0.290, slightly below the Random Forest baseline of 0.312, while substantially outperforming it after multi-geometry aggregation (0.835 versus 0.477). This transparent reporting helps clarify where the proposed method provides its main advantage.

Critical weaknesses

W1. The current evaluation is insufficient to support the central claim that learned inference can serve as a practical alternative to conventional photometric inversion.

First, no classical photometric inversion method is included in the comparison. The paper motivates the proposed approach primarily through the computational cost of iterative inversion, yet the experiments compare only with learned or statistical predictors. Consequently, the key tradeoff between inference speed and reconstruction accuracy relative to the method being replaced remains unquantified.

Second, the experimental setting does not match the sparse survey regime emphasized in the motivation. The model is evaluated using complete 72 point light curves uniformly sampled over a full rotation. No experiment evaluates sparse or irregular sampling, and no robustness analysis for photometric noise is reported. This limits the evidence for applicability to realistic survey observations.

Third, object level generalization is evaluated on only 11 unseen shapes. Although these shapes produce 1,046 light curves under different geometries, these observations share the same 11 underlying targets and should not be interpreted as 1,046 independent shape tests.

Finally, the paper defines a conventional 80:20 benchmark and states that both evaluation protocols are used, but no result for this benchmark is reported in the experimental section. The main 90:10 training and validation split is also insufficiently described to determine whether shapes are separated across the two subsets. These omissions make it difficult to assess generalization and potential object overlap.

W2. Several reported results do not fully support the paper's current interpretation of model performance.

The most important issue is the single curve result. The abstract and conclusion emphasize an R² of approximately 0.40, whereas Table 3 reports a three seed average of 0.290 with a standard deviation of 0.06 and explicitly describes 0.40 as a favorable run. The Random Forest baseline achieves 0.312 on the same single curve metric. The main empirical advantage of the CNN therefore lies in multi geometry aggregation, where it reaches 0.835 compared with 0.477 for the Random Forest, rather than in individual light curve prediction.

The training data saturation claim is also insufficiently supported. Increasing the number of training samples from 58,731 to 673,607 improves R² from 0.374 to 0.400, but this comparison appears to use single runs. The improvement is smaller than the seed variation later reported for the CNN, so repeated scaling experiments are needed before attributing the plateau primarily to physical observability.

The interpretation of the polar parameters also requires stronger evidence. Section 5.4 states that predictions for c1 and c2 tend toward the dataset mean, but the corresponding numerical R² values are not reported, and Figure 4 does not clearly demonstrate a simple collapse to a constant prediction. Moreover, the paper calls a2 the most difficult parameter while subsequently emphasizing c1 and c2 as the fundamentally weakly observable parameters. Reporting all parameter specific metrics would resolve this ambiguity.

Finally, the baseline results require additional explanation. Both the mean predictor and the MLP obtain negative R² values, but their construction and tuning are not described in sufficient detail. Negative R² is not itself inconsistent, but the current presentation makes it difficult to determine whether the comparison is adequately controlled.

Other weaknesses.

First, the model predicts only five shape parameters, while the rotation axis orientation is supplied as input metadata and the rotation period is not predicted. The manuscript should therefore distinguish more clearly between estimating asteroid shape and solving the broader shape and spin inversion problem.

Second, the stated nine dimensional geometry vector is insufficiently specified. The manuscript states that both rotation axis angles are represented using sine and cosine encoding. Together with the Sun and observer direction vectors, this description appears to require more than nine values unless some components are omitted. The exact input vector should be listed explicitly.

Third, the reported dataset units appear inconsistent. Section 3 states that the complete dataset contains more than 10 million rows and that each observational series contains 5,112 raw rows, whereas Section 5 reports 673,607 training samples. These quantities cannot refer to the same corpus under the stated definition of an observational series and should be reconciled.

Fourth, the paper emphasizes that predicted Cellinoid parameters can regenerate synthetic light curves, but no forward consistency evaluation is provided. Comparing regenerated light curves with the input observations would provide a useful physically meaningful validation of the predicted parameters.

Fifth, the abstract describes the result as significantly above a baseline, but no statistical test or confidence interval supporting this wording is reported.

Sixth, the AI Use Statement is internally inconsistent. It states that generative AI was not used for research methodology, experiments, or implementation, but later reports its use for code, experimental parameters, and modeling ideas. The statement should be clarified.

Finally, the figures could communicate the physical setup more clearly. Figure 1 does not clearly expose the asymmetric eight octant structure of the Cellinoid, while Figure 2 does not provide enough shape and observing geometry information to interpret differences among the displayed light curves.

Questions for authors
  1. Is the rotation pole assumed to be known at inference time, and does the model predict the rotation period? Since the pole orientation is provided as input, please clarify how this setting relates to the full photometric inversion problem motivated in the paper. It would also be useful to evaluate performance when the pole is estimated rather than supplied as ground truth.
  2. Are the synthetic light curves noiseless? Please specify the scattering model and photometric uncertainty model used for data generation, and evaluate the method under realistic observational noise and sparse, irregular sampling. In particular, where are the sparse sampling experiments stated as part of the paper's scope?
  3. Please report the results for the Standard Benchmark defined in Section 5.1, together with the parameter specific R squared values for b2, c1, and c2. Please also define exactly how the global R squared is computed. These results are important for evaluating the claim that c1 and c2 are poorly constrained and tend toward the dataset mean.
  4. What is the exact composition of the nine dimensional geometry metadata vector? The manuscript states that both rotation axis angles are encoded using sine and cosine, while also including Sun and observer direction vectors. Please explicitly list all input dimensions. Please also report the number of distinct training shapes and the number of observing geometries generated for each shape.
  5. How are the negative R squared values of the mean baseline and MLP with Fourier features explained? Please describe how the baselines were constructed and tuned, and report whether they received a comparable tuning budget. It would also be helpful to compare the target parameter distributions of the training set and the 11 shape test set.
  6. Table 3 reports a mean single curve R squared of 0.290 with a standard deviation of 0.06 for the CNN, compared with 0.312 for the Random Forest. On what evidence does the paper conclude that end to end deep learning is preferable for single curve prediction? Please also provide uncertainty estimates for the aggregated R squared of 0.835 and clarify whether the training, validation, and standard benchmark splits are separated by asteroid shape.
  7. The substantial improvement from multi geometry aggregation suggests that prediction errors may contain systematic object or geometry dependent components. Could the authors quantify the within shape and between shape error components and clarify whether the remaining error is primarily attributable to normalization, viewing geometry, or shape specific bias? Such an analysis would also help determine whether joint multi view inference would be more effective than modifying the normalization strategy.
  8. Could the authors report forward consistency results by comparing light curves regenerated from the predicted Cellinoids with the corresponding input light curves? A comparison with classical convex or Cellinoid inversion on the same synthetic data would also provide a direct assessment of the accuracy and computational tradeoff motivating the proposed approach.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.