Physics-constrained differentiable causal discovery of constitutive mechanisms through boundary value problems

arXiv:2611.16468 · math-ph, math.MP · Submitted 2026-11-12 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Mathematical Physics. The day in mathematical physics: rigorous models of physical systems.

Kai: We're looking at the paper: "Physics-constrained differentiable causal discovery of constitutive mechanisms through boundary value problems".

Mira: Here's the gist of the paper. Data-driven discovery of constitutive laws selects model terms that reproduce measured responses When environmental variables such as temperature and moisture co-vary in the data, a well-fitting model can still attribute an effect to the wrong variable and fail under new conditions Constitutive Models as Structural Causal Models The paper recasts constitutive model discovery as causal discovery of mechanism parent sets, where a material with internal variables is represented as a time-unrolled structural causal model Its balance laws and constitutive skeleton are hard-coded; its environment–mechanism and state–state dependencies are learnable edges.

Kai: The structural assignments for the material at each time step involve several equations defining how variables like stress, internal states, and damage evolve Feedback between internal variables is represented across time steps.

Mira: A Differentiable Scoring Mechanism Candidate graphs are scored through the boundary value problem by a weakform equilibrium gap, evaluated by strain-driven integration along full-field displacements, so neither stresses nor internal variables need to be measured.

Kai: The Gauss–Newton matrix of the score, obtained by automatic differentiation, serves as a local identifiability certificate and drives a greedy D-optimal test design.

Paper summary: Kai: So, we're looking at this paper by Michael Kraus and his team called "Physics-constrained differentiable causal discovery of constitutive mechanisms through boundary value problems." It’s about figuring out the underlying physics that governs how materials behave when you look at measured data.

Mira: Exactly. The main thesis here is that when you try to discover these constitutive laws just by fitting a model to some measured responses, there's this big problem where if your environment variables like temperature and moisture move together in the data, your model might wrongly think the effect is coming from one variable instead of the other.

Lev: So what they claim is that they recast this whole discovery process as causal discovery of mechanism parent sets. They treat a material with internal variables as a time-unrolled structural causal model where the balance laws and the constitutive skeleton are already hard-coded, and only the environment to mechanism and state to state dependencies are learnable edges.

Kai: Right, so instead of just fitting parameters, they’re learning which physical relationships are causing the material changes over time. They call these learnable edges things like how stress or internal states evolve with time based on other variables in the system.

Mira: And they score candidate graphs using a boundary value problem method—specifically a weak form equilibrium gap evaluated by strain-driven integration along full-field displacements—which is clever because it means you don't even need to measure stresses or those internal variables directly to get the score.

Lev: That's the core of their methodology. They use automatic differentiation to get a Gauss–Newton matrix from that score, and this matrix acts like a local identifiability certificate, which then drives a greedy D-optimal test design for selecting the best candidate graphs without knowing the true model beforehand.

Kai: It’s essentially using physics constraints—the boundary value problem—to guide the learning process and pick the right edges in that structural causal model, which is pretty cool because it links experimental design directly to causal inference.

Mira: And they even treat climate-chamber protocols as interventions, which they interpret as setting temperature and moisture independently, effectively removing the direct causal link from the environment variables to those specific variables.

Lev: The paper notes that this Gauss–Newton matrix is nearly singular along directions where your observational protocols are collinear, and this certificate predicts attribution errors caused by those collinear parents but doesn't predict the bias introduced by measurement noise in that weak-form score.

Kai: That distinction is important because it tells us what the method can and can’t handle. It shows them where their current scoring mechanism breaks down when there's noise involved, which leads right into their experimental design part.

Paper summary: Mira: The experimental design part is where they show the results on a synthetic study involving a perforated plate. They found that using these climate-chamber interventions, they recovered the true graph exactly in seven of ten runs when the displacement noise was about zero point zero one pixel, and crucially, there was no attribution error at all <ref:2611.16468#pg2>.

Lev: That's a strong result for their method because it shows that intervention data can recover the truth very accurately under controlled conditions, which is what we need to see for real hardware testing.

Kai: But they also showed what happens when you only use observational data correlated with temperature and moisture—with a correlation of negative zero point nine nine eight—that never recovers the true graph.

Mira: So, the paper demonstrates that while fitting models to data is easy, if your environment variables are coupled in the data, you can end up attributing effects to the wrong variable and failing when you test under new conditions.

Lev: The limitation they point out is that their Gauss–Newton certificate predicts those attribution errors from collinear parents, but it doesn't predict the noise induced bias of the weak-form score itself. They found that a measurement noise of zero point zero three pixel biases the viscous branch enough to create false edges under any protocol.

Kai: That’s a practical limitation for me because on real superconducting qubits or anything requiring high precision, you can't ignore measurement noise; it can actually create spurious physical connections in your model.

Mira: They also mentioned that the acyclicity of the instantaneous graph means the material update is explicit in its order of evaluation, and this removed a spurious algebraic loop that two of five unconstrained runs still kept. That’s good for keeping the math clean.

Lev: The paper concludes that their greedy D-optimal design selected test sets ranked second or third out of one hundred twenty-six candidates when they started from observational data alone, even with a consistent normalization.

Kai: So what this means for us is that if we're just looking at observational data, the best tests we can design are mediocre, ranking only second or third among all possibilities.

Mira: The paper suggests that if you use climate-chamber interventions to set temperature and moisture independently, you can recover the true mechanism graph exactly seven out of ten times without any attribution error. That’s a powerful tool for disentangling coupled environmental effects.

Lev: For someone building real hardware, this means that designing experiments based on physical intervention protocols can actually yield results closer to the truth than just relying on what you observe naturally.

Paper summary: Kai: So it seems the paper is showing how to use physics constraints, through these boundary value problems, to turn experimental design into a causal inference problem for constitutive models.

Mira: That’s right. The core idea is that you can use those computable certificates—like the Gauss–Newton matrix—to guide your experiments so you don't get stuck with poor observational results that mislead you about what's actually happening in the material.

Lev: But they are clear about their own limitation too: this certificate predicts attribution errors from collinear parents, but it doesn't predict the bias caused by measurement noise in that weak-form score, which they say remains a main limitation of their current approach.

Kai: So the next step for them is to figure out how to get a bias-aware score, maybe using output-error formulations or implicit differentiation based on the measurement model Fisher information.

Mira: I think it’s about making the scoring mechanism robust against noise, which is crucial if we are trying to apply this kind of discovery to things that have inherent measurement uncertainty.

Lev: If you were running this on real hardware, you'd need to account for that noise bias because as they showed, even small amounts of noise can introduce false edges in the model structure.

Kai: It’s a big step from just fitting curves to figuring out the underlying causal structure of how things change over time under varying physical conditions.

Mira: And the conclusion is that climate-chamber interventions do recover the true mechanism graph exactly seven out of ten times and make no attribution error in any run, which is a solid performance metric for their method.

Lev: But remember, it still relies on that computable certificate to predict attribution errors from collinear parents, and they haven't solved the noise bias problem yet.

Kai: So if we are trying to understand complex material behavior driven by coupled environmental factors, this framework gives us a way to test hypotheses much more intelligently than just collecting raw data.

Mira: It moves the focus from just finding a good fit to identifying the causal structure that causes that fit, which is where the real physical understanding lies.

Lev: For people working on experimental setups, it suggests that controlled interventions are far more informative for structural discovery than passive observational data when environmental variables are intertwined.

Kai: So we've seen how they use this to test models and find better designs, but the next challenge is making that design process totally immune to the noise floor of the measurement itself.

Conclusion: Kai: So, we're looking at this paper, "Physics-constrained differentiable causal discovery of constitutive mechanisms through boundary value problems." It’s about using physics to figure out the underlying rules of how materials change over time when you look at measured data.

Mira: Exactly. The core idea is that instead of just trying to fit a curve and guess what caused it, they treat this like finding the causal structure—the actual mechanism—that makes the material behave that way.

Lev: It frames the material as a structural causal model where the basic laws are already fixed, and only certain connections between variables are things we learn.

Kai: They use this boundary value problem approach to score candidate structures, which means they don't even need to measure all those internal stresses or states directly to get a good estimate of the structure.

Mira: They get this certificate, called the Gauss–Newton matrix from automatic differentiation, which tells you how good a structure is locally.

Lev: And that certificate then drives an experimental design process, turning finding the right model into designing the right experiments to test it.

Kai: The results are pretty striking on their synthetic study with a perforated plate. They found that if they used climate-chamber interventions—setting temperature and moisture separately—they recovered the true mechanism graph exactly seven out of ten times.

Mira: And crucially, in those successful runs, there was zero attribution error regarding the temperature or moisture effects. That's what they were aiming for with this method.

Lev: That level of recovery under controlled intervention is what would be needed to run this on real hardware, though we have to keep an eye on the noise.

Kai: But they also showed that just looking at observational data correlated with temperature and moisture didn't work well at all; it never recovered the true graph.

Mira: And they flagged a specific limitation: while the method predicts errors from things like two variables being dependent, it doesn't predict the bias caused by measurement noise itself.

Lev: They found that even small amounts of pixel noise can introduce false connections in their model structure, which is something we need to worry about when translating this to actual experimental setups.

Kai: So essentially, this paper shows how you can use physical constraints and smart test design to get much more reliable material models than just fitting data.

Mira: It moves the focus from finding a good fit to figuring out the specific causal structure causing that fit, which is where we need to be for real physics.

Lev: It suggests that controlled interventions are way more informative than passive measurements when those environmental variables are mixed up in the data.

Kai: So what this means for us is that if you're trying to understand complex material behavior driven by coupled factors, you can design experiments much smarter than just collecting raw data points.

Mira: But the next big challenge they point out is figuring out how to get a score that isn't biased by the measurement noise itself.

Lev: That means they need bias-aware scores, maybe using output-error formulations or looking at the measurement model Fisher information through implicit differentiation.

Kai: So if we're trying to make this work on real superconducting qubits or anything requiring high fidelity, accounting for that noise floor is a massive step forward.

Michael Kraus

Institute of Structural Mechanics and Design (ISM+D), Technical University of Darmstadt

math-ph, math.MP

Submitted: 2026-11-12

Updated: 2026-11-12

Code: https://github.com/ax-ml/jax

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 91/100

The gist: Data-driven discovery of constitutive laws selects model terms that reproduce measured responses When environmental variables such as temperature and moisture co-vary in the data…

Key concepts

Constitutive Models as Structural Causal Models
This concept treats a material's internal state (like stress or damage) as a system whose evolution is governed by fixed physical laws (balance laws). The paper frames this as finding the causal structure—which variables influence which others—by learning the variable-to-variable dependencies that are not explicitly defined in the basic physics equations.
Differentiable Scoring Mechanism
Instead of relying on traditional measurements, candidate models are scored using a mathematical problem called a weakform equilibrium gap. This score is calculated by integrating strain along full field displacements, allowing the use of automatic differentiation to efficiently test thousands of potential connections without needing to measure internal stresses or variables directly.
Intervention Design and Identifiability
Climate-chamber protocols are used as controlled 'interventions'—meaning researchers deliberately set temperature (T) and moisture (w) independently. The paper uses a mathematical certificate derived from the scoring process to predict exactly where attribution errors will occur when using observational data, highlighting the difference between controlled and uncontrolled testing.

Terminology

Summary

Data-driven discovery of constitutive laws selects model terms that reproduce measured responses When environmental variables such as temperature and moisture co-vary in the data, a well-fitting model can still attribute an effect to the wrong variable and fail under new conditions

Constitutive Models as Structural Causal Models

The paper recasts constitutive model discovery as causal discovery of mechanism parent sets, where a material with internal variables is represented as a time-unrolled structural causal model Its balance laws and constitutive skeleton are hard-coded; its environment–mechanism and state–state dependencies are learnable edges. The structural assignments for the material at each time step involve several equations defining how variables like stress, internal states, and damage evolve Feedback between internal variables is represented across time steps. p. 6

Where the paper says it

The variables at a material point and time tn+1 are grouped as follows: • environment z = (T , ˆ wˆ): normalised temperature Tˆ = (T − T0)/20 K and moisture wˆ = (w − w0)/∆wref; • kinematics: total strain ε, obtained from displacements; • mechanism parameters m = (E∞, E1, τ, κ0) and eigenstrain εθ; • latent internal states q = (εv, κ, D): viscous strain, damage history variable and damage; • stress σ, which is never observed. The structural assignments of one time step are, in their order of evaluation: mk := m0 k exp θkT Tˆ + θkw wˆ , k ∈ {E∞, E1, κ0}, (2a) εθ := 10−3 θθT Tˆ + θθw wˆ (1, 1, 0)T , εm := ε − εθ, (2b) τ := τ 0 exp θτT Tˆ + θτw wˆ + θ lag τD D n + θ inst τD D n+1 , (2c) ε n+1 v := ε n v + (∆t/τ ) εm 1 + ∆t/τ , (2d) εd := εm − θDv ε n+1 v , κn+1 := max κ n , ∥εd∥ , (2e) D n+1 := Dmax 1 − exp − ⟨κ n+1 − κ0⟩/κf , (2f) σ := (1 − D n+1) C¯ h E∞ εm + E1(εm − ε n+1 v ) i .

A Differentiable Scoring Mechanism

Candidate graphs are scored through the boundary value problem by a weakform equilibrium gap, evaluated by strain-driven integration along full-field displacements, so neither stresses nor internal variables need to be measured. The Gauss–Newton matrix of the score, obtained by automatic differentiation, serves as a local identifiability certificate and drives a greedy D-optimal test design. p. 1

Where the paper says it

Climate-chamber protocols are treated as interventions. The Gauss–Newton matrix of the score, obtained by automatic differentiation, serves as a local identifiability certificate and drives a greedy D-optimal test design.

This approach allows for the scoring of candidate graphs using differentiable finite element discretisation. p. 3
Where the paper says it

We treat a constitutive model as a hybrid SCM whose structural equations are partly fixed by balance laws and partly learned, and we score candidate graphs through the BVP with a differentiable finite element discretisation. The approach rests on four contributions: 1.

Intervention Design and Identifiability

Climate-chamber protocols are treated as interventions, which are interpreted as setting T and w independently, removing the edges U → T and U → w. The Gauss–Newton matrix of the score serves as a local identifiability certificate and is nearly singular along the collinear directions of observational protocols, which are characterized by Proposition 2. p. 3

Where the paper says it

The Gauss–Newton matrix of the score, computed by automatic differentiation, serves as a local identifiability certificate. It is nearly singular along the collinear directions of observational protocols and well conditioned for interventional ones.

This certificate predicts attribution errors but not the noise-induced bias of the weak-form score. p. 2
Where the paper says it

5 %; interventionally discovered models stay within 3. A correlation sweep shows that the certificate predicts attribution errors but not the noiseinduced bias of the weak-form score, which remains the main limitation.

Experimental Design and Performance

The Gauss–Newton matrix drives a greedy D-optimal selection of tests from a candidate library, turning causal identifiability into an experimental design problem. In a synthetic study on a perforated plate, interventional data recover the true graph exactly in 7 of 10 runs at a displacement noise of about 0. p. 1

Where the paper says it

The Gauss–Newton matrix of the score, obtained by automatic differentiation, serves as a local identifiability certificate and drives a greedy D-optimal test design. In a synthetic study on a perforated plate of a hygro-thermo-viscoelastic damaging material (8 true edges among 13 candidates), interventional data recover the true graph exactly in 7 of 10 runs at a displacement noise of about 0.

01 px and make no attribution error. p. 1
Where the paper says it

In a synthetic study on a perforated plate of a hygro-thermo-viscoelastic damaging material (8 true edges among 13 candidates), interventional data recover the true graph exactly in 7 of 10 runs at a displacement noise of about 0. 01 px and make no attribution error.

Observational data with a temperature–moisture correlation of −0. p. 1
Where the paper says it

01 px and make no attribution error. Observational data with a temperature–moisture correlation of −0.

998 never recover it. p. 1
Where the paper says it

Observational data with a temperature–moisture correlation of −0. 998 never recover it.

Limitations and Future Directions

The Gauss–Newton certificate predicts the attribution errors caused by collinear parents, but it does not predict the bias that measurement noise induces in the weak-form score. The study shows that noise of 0. p. 3

Where the paper says it

It is nearly singular along the collinear directions of observational protocols and well conditioned for interventional ones. A correlation sweep shows what it captures (attribution errors) and what it misses (noise-induced bias).

03 px biases the viscous branch enough to create false edges under any protocol. p. 6
Where the paper says it

The viscous update is a backward-Euler step of ε˙v = (εm − εv)/τ . Two kinds of edges are distinguished.

Acyclicity of the instantaneous graph makes the material update explicit in its order of evaluation, and it removed a spurious algebraic loop that two of five unconstrained runs retained. p. 3
Where the paper says it

Feedback between internal variables is represented across time steps. Acyclicity of the instantaneous subgraph guarantees an explicit order of evaluation of the material update.

The next steps are bias-aware scores, either output-error formulations or the measurement-model Fisher information by implicit differentiation. p. 9
Where the paper says it

F is a property of the score and is used as a local identifiability certificate. It is not the Fisher information of the measurement model: displacement noise enters the residual nonlinearly, through strains and the history recursion, so J TJ yields no Cramér–Rao bound.

Conclusions

The main findings of the synthetic study are that climate-chamber interventions recover the true mechanism graph exactly in 7 of 10 runs and make no T/w attribution error in any run. The attribution errors are predicted by a computable certificate. p. 2

Where the paper says it

5 %; interventionally discovered models stay within 3. A correlation sweep shows that the certificate predicts attribution errors but not the noiseinduced bias of the weak-form score, which remains the main limitation.

Observationally discovered models predict a climate on their training manifold within 3. p. 2
Where the paper says it

confounded directions, their score is 250–380 times flatter. Observationally discovered models reproduce a climate within their training range to 3.

6 % on average, but a hot and humid climate with 16. p. 2
Where the paper says it

6 % error but mispredict a hot and humid climate by 16. 9 % on average and up to 41.

9 % mean and 41.5 % maximum error. p. 1
Where the paper says it

In a synthetic study on a perforated plate of a hygro-thermo-viscoelastic damaging material (8 true edges among 13 candidates), interventional data recover the true graph exactly in 7 of 10 runs at a displacement noise of about 0. 01 px and make no attribution error.

The greedy D-optimal design from observational data alone selected test sets ranked second or third of 126 under a consistent normalisation, for three different starting models. p. 2
Where the paper says it

The D-optimal design, computed without knowledge of the truth, selects test sets ranked second or third of 126 candidates. Keywords: constitutive model discovery, causal inference, structural causal models, differentiable physics, identifiability, optimal experimental design 1.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Adaptive causal discovery of constitutive mechanisms can select model terms by minimizing a score derived from strain-driven integration along full-field displacements, which requires only measured displacement fields and reaction forces rather than stresses or internal variables.

  2. The system can perform climate chamber protocol interventions to recover the true mechanism graph exactly in 7 of 10 runs at a typical noise level of about 0.01 px and make no attribution error, whereas observational data with a temperature moisture correlation of -0.998 never recover it.

  3. A Fisher-based design can select near-optimal tests without prior knowledge of the true graph by using the Gauss–Newton matrix of the score as a local identifiability certificate to drive a greedy D-optimal test design.

  4. The system can distinguish between observational and interventional protocols by observing that Along the collinear directions of Proposition 2, the score of the observational protocol is 250–380 times flatter than that of the interventional protocol.

  5. The system can identify and mitigate noise-induced bias by recognizing that Noise biases the weak-form score, and this bias caused a spurious viscous edge and the loss of a weak damage edge in all observational-type designs, but rarely under the climate-chamber design.

  6. The system can perform predictive simulations for new climates with high fidelity; models discovered from interventional data stay within 3.2 % error, compared to 16.9 % mean and up to 41.5 % maximum error for observational models in hot and humid conditions.

Abstract

Data-driven constitutive model discovery is recast as causal discovery of mechanism parent sets using time-unrolled structural causal models. Hard-coding balance laws and constitutive skeletons, candidate graphs score environment--mechanism and state--state edges via weak-form equilibrium gaps evaluated from full-field displacements, avoiding the need to measure stresses or internal variables. An adaptive lasso and a NOTEARS-type constraint ensure graph acyclicity and an explicit evaluation order, treating climate-chamber protocols as interventions. Automatic differentiation yields the score's Gauss--Newton matrix as a local identifiability certificate, driving greedy D-optimal test design. In a synthetic study on a hygro-thermo-viscoelastic damaging plate (8 true edges among 13 candidates), interventional data recover the exact true graph in 7 of 10 runs (0.01px displacement noise) with zero attribution error. Conversely, observational data with a temperature--moisture correlation of-0.998 fail entirely because confounded scores are 250--380 times flatter. Observationally discovered models reproduce training climates (3.6% error) but mispredict hot and humid climates by up to 41.5% (16.9% average), whereas interventionally discovered models stay within 3.2%. A correlation sweep shows the certificate successfully predicts attribution errors, though weak-form score noise bias remains a primary limitation. Computed without knowledge of the truth, the D-optimal design selects test sets ranked second or third out of 126 candidates.

Sources

Related papers