pith. sign in

arxiv: 2605.07467 · v1 · submitted 2026-05-08 · 💻 cs.LG · cs.AI· cs.ET

Physical Simulators as Do-Operators: Causal Discovery under Latent Confounders for AI-for-Science

Tsuyoshi Okita This is my paper

Pith reviewed 2026-05-11 01:47 UTC · model grok-4.3

classification 💻 cs.LG cs.AIcs.ET
keywords causal discoverylatent confoundersphysical simulatorsdo-operatorsAI for scienceinterventional datamolecular design
0
0 comments X

The pith

Physical simulators serve as do-operators to discover causal structures under latent confounders with O(d) single-variable interventions.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces CFM-SD, which uses first-principles physical simulators as do-operators in Pearl's interventional calculus. This allows causal discovery even when latent confounders are present, a common issue in AI-for-Science applications such as molecular design. Theoretically, it shows that the causal structure for d variables is identifiable using only O(d) interventions, the minimum possible under physical constraints. In practice, it outperforms baselines on synthetic data with higher F1 scores and achieves substantial bias reduction on real scientific datasets.

Core claim

By treating physical simulators directly as do-operators, CFM-SD enables identification of causal structures in the presence of latent confounders. The method proves that d-variable causal graphs are identifiable with O(d) single-variable interventions. Evaluations show average F1 of 0.800 on synthetic data with varying gamma and 57-58% bias reduction in real tasks like toxicity prediction.

What carries the argument

CFM-SD (Causal Flow Matching with Simulation Data) which integrates physical simulators as interventions to handle latent confounders in causal structure learning.

If this is right

  • Causal structures become identifiable under latent confounders using the minimal number of physical interventions.
  • Downstream scientific predictions suffer less bias, as seen in 57-58% reduction in molecular and battery tasks.
  • Existing methods assuming causal sufficiency can be outperformed in realistic settings.
  • The approach provides a bridge between synthetic interventional methods and real physical data.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If simulators have inaccuracies, additional modeling of simulator error might further improve results.
  • This method could be extended to choose optimal interventions adaptively to minimize the number needed.
  • Applications in other fields with expensive physical experiments may benefit similarly.

Load-bearing premise

That first-principles physical simulators can be used directly as do-operators without modeling their own inaccuracies or how they interact with latent confounders.

What would settle it

Observing whether CFM-SD maintains its performance advantage when applied to a simulator with known significant discrepancies from real physics, compared to baselines that explicitly model uncertainty.

Figures

Figures reproduced from arXiv: 2605.07467 by Tsuyoshi Okita.

Figure 1
Figure 1. Figure 1: Overall architecture of CFM-SD. Phase 1: [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
read the original abstract

Existing interventional causal discovery methods -- IGSP, DCDI, ENCO -- assume causal sufficiency (no latent confounders) and rely on virtual interventions in synthetic simulators. In AI-for-Science settings such as molecular design and materials science, latent confounders are ubiquitous and real interventions (e.g., physics-based simulations) require hours to days per data point. We propose CFM-SD (Causal Flow Matching with Simulation Data), which uses first-principles physical simulators as do-operators in Pearl's interventional calculus to simultaneously handle latent confounders and real interventional data. Theoretically, $d$-variable causal structure is identifiable with $O(d)$ single-variable interventions -- the minimum under physical realizability constraints. In Intrinsic Evaluation on synthetic data ($\gamma=0.2$--$0.8$), CFM-SD achieves average F1$=0.800$ vs. F1$=0.127$--$0.562$ for all baselines. In Extrinsic Evaluation on real scientific data, CFM-SD achieves 57--58\% bias reduction in molecular toxicity prediction and battery electrolyte optimization, demonstrating practical value beyond synthetic benchmarks.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes CFM-SD, which treats first-principles physical simulators as exact do-operators within Pearl's interventional calculus to perform causal discovery under latent confounders. It claims that d-variable causal structures are identifiable from O(d) single-variable interventions (the minimum under physical realizability), reports average F1=0.800 on synthetic data with latent confounders (γ=0.2–0.8) versus 0.127–0.562 for baselines, and shows 57–58% bias reduction on real molecular toxicity and battery electrolyte tasks.

Significance. If the identifiability result and simulator-as-do-operator assumption hold, the work offers a practical route to causal discovery in AI-for-Science domains where latent confounders are common and interventions are costly. The explicit use of physical simulators rather than purely synthetic interventions, together with both intrinsic and extrinsic evaluations, is a concrete strength that could influence downstream modeling in chemistry and materials.

major comments (2)
  1. [§3] §3 (Identifiability Theorem): The central claim that d-variable structure is identifiable with O(d) interventions rests on treating the physical simulator as a faithful implementation of P(· | do(X=x)) that is unaffected by its own numerical/discretization errors or by latent confounders outside the observed variables. No quantitative bound or sensitivity analysis is provided for simulator mismatch, which directly undermines the theorem and the reported F1/bias numbers.
  2. [§4.2] §4.2 (Intrinsic Evaluation): The synthetic data generation with γ=0.2–0.8 is described as incorporating latent confounders, yet the precise mechanism by which the simulator is queried under do-interventions (and how baselines are modified to use the same simulator) is not specified. Without this protocol, it is impossible to verify whether the F1 gap is attributable to CFM-SD or to differences in intervention access.
minor comments (2)
  1. [Abstract] Abstract and §1: The phrasing 'real interventional data' is used for simulator outputs; clarify that these are simulated interventions and distinguish them from physical lab experiments.
  2. [Notation] Notation: The definition of the causal flow matching objective and its relation to the do-calculus should be stated explicitly with a single consistent set of symbols rather than scattered across sections.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the thoughtful and constructive report. The two major comments highlight important points about theoretical assumptions and experimental clarity. We address each below and will revise the manuscript to incorporate additional analysis and details.

read point-by-point responses

  1. Referee: [§3] §3 (Identifiability Theorem): The central claim that d-variable structure is identifiable with O(d) interventions rests on treating the physical simulator as a faithful implementation of P(· | do(X=x)) that is unaffected by its own numerical/discretization errors or by latent confounders outside the observed variables. No quantitative bound or sensitivity analysis is provided for simulator mismatch, which directly undermines the theorem and the reported F1/bias numbers.

    Authors: We agree that the identifiability result in Theorem 1 is stated under the assumption of a perfect simulator that exactly realizes the do-operator. This is a standard modeling choice in interventional causal discovery literature when simulators are used, and the paper explicitly lists the assumption in §3. To directly address the concern about mismatch, we will add a new sensitivity analysis subsection (and corresponding appendix) in the revision. This analysis will inject controlled numerical noise and discretization errors into the simulator outputs at varying levels and report the resulting degradation in F1 scores, providing quantitative bounds on robustness. The core theorem remains valid under the stated assumptions, but the added empirical study will better contextualize the reported performance numbers. revision: yes

  2. Referee: [§4.2] §4.2 (Intrinsic Evaluation): The synthetic data generation with γ=0.2–0.8 is described as incorporating latent confounders, yet the precise mechanism by which the simulator is queried under do-interventions (and how baselines are modified to use the same simulator) is not specified. Without this protocol, it is impossible to verify whether the F1 gap is attributable to CFM-SD or to differences in intervention access.

    Authors: We acknowledge that the intervention protocol in the intrinsic evaluation was underspecified. In the revised §4.2 we will add a detailed protocol description: the simulator is queried for a do-intervention on variable X_i by fixing X_i to a chosen value while integrating the remaining physical dynamics forward in time (with latent confounders realized as additive unobserved Gaussian noise terms drawn once per trajectory and held fixed across interventions). All baselines (IGSP, DCDI, ENCO) were given identical access to these simulator queries; their implementations were adapted only to respect the single-variable intervention constraint and to use the same number of interventions O(d). This ensures the F1 comparison isolates the effect of CFM-SD's flow-matching objective rather than differences in intervention access. revision: yes

Circularity Check

0 steps flagged

No circularity; identifiability follows from Pearl calculus with external simulator assumptions

full rationale

The paper's derivation of d-variable identifiability from O(d) interventions applies standard Pearl do-calculus to first-principles simulators treated as external do-operators. No quoted equation or self-citation reduces the theorem to a fitted parameter, renamed pattern, or self-referential definition. The F1 and bias-reduction numbers are downstream evaluations on held-out data, not inputs that force the theoretical claim. This matches the default expectation that most papers are non-circular when the central result rests on independent causal theory plus external simulators.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on the domain assumption that physical simulators implement do-interventions faithfully and on standard causal identifiability results adapted to this interventional regime.

axioms (2)
  • domain assumption Physical simulators can be treated as exact do-operators in Pearl's interventional calculus
    Invoked when the method substitutes simulator runs for interventions.
  • domain assumption Causal structure remains identifiable from O(d) single-variable interventions even with latent confounders
    Core of the theoretical guarantee stated in the abstract.

pith-pipeline@v0.9.0 · 5504 in / 1291 out tokens · 57855 ms · 2026-05-11T01:47:26.975296+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

40 extracted references · 40 canonical work pages · 1 internal anchor

  1. [1]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Recovering from Selection Bias in Causal and Statistical Inference , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    work page

  2. [2]

    British Journal of Mathematical and Statistical Psychology , volume=

    Perturbation graphs, invariant causal prediction and causal relations in psychology , author=. British Journal of Mathematical and Statistical Psychology , volume=

    work page

  3. [3]

    and Burke, Kieron and Ernzerhof, Matthias , title =

    Perdew, John P. and Burke, Kieron and Ernzerhof, Matthias , title =. Physical Review Letters , volume =

    work page

  4. [4]

    and Scheines, Richard and Heckerman, David , title =

    Spirtes, Peter and Glymour, Clark N. and Scheines, Richard and Heckerman, David , title =

    work page

  5. [5]

    Journal of Machine Learning Research , volume =

    Chickering, David Maxwell , title =. Journal of Machine Learning Research , volume =

    work page

  6. [6]

    Projector Augmented-Wave Method , journal =

    Bl. Projector Augmented-Wave Method , journal =

    work page

  7. [7]

    Computational Materials Science , volume =

    Dal Corso, Andrea , title =. Computational Materials Science , volume =

    work page

  8. [8]

    Shimizu, Shohei and Hoyer, Patrik O. and Hyv. A Linear Non-. Journal of Machine Learning Research , volume =

    work page

  9. [9]

    and Xing, Eric P

    Zheng, Xun and Aragam, Bryon and Ravikumar, Pradeep K. and Xing, Eric P. , title =. Advances in Neural Information Processing Systems , volume =

    work page

  10. [10]

    Causal Inference by Using Invariant Prediction , journal =

    Peters, Jonas and B. Causal Inference by Using Invariant Prediction , journal =

    work page

  11. [11]

    Lipman, Yaron and Chen, Ricky T. Q. and Ben-Hamu, Heli and Nickel, Maximilian and Le, Matthew , title =. arXiv preprint arXiv:2210.02747 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  12. [12]

    Journal of Machine Learning Research , volume =

    Shimizu, Shohei and Inazumi, Takanori and Sogawa, Yasuhiro and Hyv. Journal of Machine Learning Research , volume =

    work page

  13. [13]

    Pearl, Judea , title =

    work page

  14. [14]

    Elements of Causal Inference: Foundations and Learning Algorithms , publisher =

    Peters, Jonas and Janzing, Dominik and Sch. Elements of Causal Inference: Foundations and Learning Algorithms , publisher =

    work page

  15. [15]

    Learning Neural Causal Models with Active Interventions , booktitle =

    Scherrer, Nino and Bilaniuk, Olexa and Annadani, Yashas and Goyal, Anirudh and Schwab, Patrick and Sch. Learning Neural Causal Models with Active Interventions , booktitle =

    work page

  16. [16]

    Interventions, Where and How?

    Tigas, Panagiotis and Annadani, Yashas and Jesson, Andrew and Sch. Interventions, Where and How?. Advances in Neural Information Processing Systems , volume =

    work page

  17. [17]

    Toward Causal Representation Learning , journal =

    Sch. Toward Causal Representation Learning , journal =

    work page

  18. [18]

    , title =

    Raissi, Maziar and Perdikaris, Paris and Karniadakis, George E. , title =. Journal of Computational Physics , volume =

    work page

  19. [19]

    Molecular Informatics , volume =

    Tropsha, Alexander , title =. Molecular Informatics , volume =

    work page

  20. [20]

    and Nagy, Stacey R

    Denison, Michael S. and Nagy, Stacey R. , title =. Annual Review of Pharmacology and Toxicology , volume =

    work page

  21. [21]

    and Rossoshek, Anna and Simeonov, Anton , title =

    Huang, Ruili and Xia, Menghang and Nguyen, Dac-Trung and Zhao, Tongan and Sakamuru, Srilatha and Zhao, Jinghua and Shahane, Sampada A. and Rossoshek, Anna and Simeonov, Anton , title =. Frontiers in Environmental Science , volume =

    work page

  22. [22]

    The Econometrics Journal , volume =

    Chernozhukov, Victor and Chetverikov, Denis and Demirer, Mert and Duflo, Esther and Hansen, Christian and Newey, Whitney and Robins, James , title =. The Econometrics Journal , volume =

    work page

  23. [23]

    and Ding, Peng , title =

    VanderWeele, Tyler J. and Ding, Peng , title =. Annals of Internal Medicine , volume =

    work page

  24. [24]

    Structural Intervention Distance for Evaluating Causal Graphs , journal =

    Peters, Jonas and B. Structural Intervention Distance for Evaluating Causal Graphs , journal =

    work page

  25. [25]

    Physical Review , volume =

    Kohn, Walter and Sham, Lu Jeu , title =. Physical Review , volume =

    work page

  26. [26]

    and Cococcioni, Matteo and Dabo, Ismaila and others , title =

    Giannozzi, Paolo and Baroni, Stefano and Bonini, Nicola and Calandra, Matteo and Car, Roberto and Cavazzoni, Carlo and Ceresoli, Davide and Chiarotti, Guido L. and Cococcioni, Matteo and Dabo, Ismaila and others , title =. Journal of Physics: Condensed Matter , volume =

    work page

  27. [27]

    and Ceder, Gerbrand , title =

    Deng, Bowen and Zhong, Peichen and Jun, KyuJung and Riebesell, Janosh and Han, Kevin and Bartel, Christopher J. and Ceder, Gerbrand , title =. Nature Machine Intelligence , volume =

    work page

  28. [28]

    Energy Storage Materials , volume =

    Su, Chi-Cheung and He, Meinan and Amine, Rachid and Chen, Zonghai and Sahore, Ritu and Dietz Rago, Nancy and Amine, Khalil , title =. Energy Storage Materials , volume =

    work page

  29. [29]

    Advances in Neural Information Processing Systems , year=

    Permutation-based Causal Inference Algorithms with Interventions , author=. Advances in Neural Information Processing Systems , year=

    work page

  30. [30]

    Conference on Uncertainty in Artificial Intelligence , year=

    Permutation-Based Causal Structure Learning with Unknown Intervention Targets , author=. Conference on Uncertainty in Artificial Intelligence , year=

    work page

  31. [31]

    Journal of Machine Learning Research , volume=

    Joint Causal Inference from Multiple Contexts , author=. Journal of Machine Learning Research , volume=

    work page

  32. [32]

    Advances in Neural Information Processing Systems , year=

    Differentiable Causal Discovery from Interventional Data , author=. Advances in Neural Information Processing Systems , year=

    work page

  33. [33]

    International Conference on Learning Representations , year=

    Efficient Neural Causal Discovery without Acyclicity Constraints , author=. International Conference on Learning Representations , year=

    work page

  34. [34]

    International Conference on Machine Learning , year=

    Bicycle: Bidirectional Causal Discovery with Interventions , author=. International Conference on Machine Learning , year=

    work page

  35. [35]

    Advances in Neural Information Processing Systems , year=

    Experimental Design for Learning Causal Graphs with Latent Variables , author=. Advances in Neural Information Processing Systems , year=

    work page

  36. [36]

    Characterization and greedy learning of interventional

    Hauser, Alain and B. Characterization and greedy learning of interventional. Journal of Machine Learning Research , volume=

    work page

  37. [37]

    Advances in Neural Information Processing Systems , year=

    Almod. Advances in Neural Information Processing Systems , year=

    work page

  38. [38]

    Advances in Neural Information Processing Systems , year=

    Causal Normalizing Flows: from Theory to Practice , author=. Advances in Neural Information Processing Systems , year=

    work page

  39. [39]

    Journal of Machine Learning Research , volume=

    A kernel two-sample test , author=. Journal of Machine Learning Research , volume=

    work page

  40. [40]

    Annals of Statistics , volume=

    Statistical guarantees for the EM algorithm: From population to sample-based analysis , author=. Annals of Statistics , volume=

    work page