Differentiable graph neural network simulator for forward and inverse modeling of multi-layered slope system with multiple material properties

Chenying Liu; Jorge Macedo; Yongjin Choi

arxiv: 2504.15938 · v3 · submitted 2025-04-22 · ⚛️ physics.geo-ph

Differentiable graph neural network simulator for forward and inverse modeling of multi-layered slope system with multiple material properties

Yongjin Choi , Jorge Macedo , Chenying Liu This is my paper

Pith reviewed 2026-05-22 18:51 UTC · model grok-4.3

classification ⚛️ physics.geo-ph

keywords graph neural network simulatorgranular flowsmulti-layered slopesinverse modelingdifferentiable simulationmaterial property back-calculationslope runoutdam failure

0 comments

The pith

A differentiable graph neural network simulator enables efficient forward modeling of multi-layered slope runouts and inverse back-calculation of material properties.

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

The paper presents a framework that uses a fine-tuned graph neural network to simulate granular flows in multi-layered slope systems, accounting for both friction angle and cohesion to handle realistic heterogeneous materials. This forward simulator replicates the behavior of traditional material point method calculations but with substantial speed improvements. The differentiability of the model then allows it to be embedded in an optimization loop that infers the material strengths responsible for a given runout geometry in a dam system example. Results from the inverse step are obtained within minutes and align well with the target values, suggesting the approach can support practical engineering tasks in slope stability analysis.

Core claim

The authors show that a graph neural network simulator, trained on column collapse and multi-layered slope data with multiple material properties, can be made differentiable to support both rapid forward simulation of runout and gradient-based inverse inference of the friction and cohesion values that produce an observed failure runout in complex slope systems.

What carries the argument

Differentiable graph neural network (GNS) that takes friction angle and cohesion as inputs for each material layer and uses reverse-mode automatic differentiation combined with L-BFGS-B optimization for the inverse component.

If this is right

Multi-layered slope runouts can be simulated orders of magnitude faster than with conventional MPM while preserving multi-material flow features.
Material property inference from observed runout becomes feasible in minutes rather than requiring exhaustive trial simulations.
The framework directly supports engineering assessments of realistic slope systems such as dams with heterogeneous fill materials.
Forward and inverse tasks can be performed within a unified differentiable modeling pipeline.

Where Pith is reading between the lines

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

Similar differentiable simulators could be applied to other geotechnical inverse problems involving heterogeneous media and observed displacement patterns.
The current training on limited cases suggests testing generalization on additional slope configurations to confirm robustness for broader use.
Integration with field monitoring data might allow updating strength estimates in near real time as new runout observations become available.

Load-bearing premise

The fine-tuned GNS trained on column collapse and selected multi-layered cases is assumed to capture the essential coupled flow dynamics and boundary conditions sufficiently well for the dam system inverse problem.

What would settle it

A direct comparison where the material properties inferred by the GNS optimization are used to drive a full MPM simulation and the resulting runout is checked for close agreement with the target geometry used in the inverse step.

read the original abstract

Graph neural network simulators (GNS) have emerged as a computationally efficient tool for simulating granular flows. Previous efforts have been limited to simplified homogeneous geometries characterized only by the friction angle, which does not reflect the complexity of realistic slopes encountered in engineering practice. This study introduces a differentiable GNS framework designed for multi-layered slope systems comprising both forward and inverse modeling components. The forward component relies on a fine-tuned GNS that incorporates both friction angle and cohesion. Its performance is demonstrated through column collapse and multi-layered slope runout simulations, where the GNS replicates multi-material flow dynamics while achieving significant computational speedup over the Material Point Method (MPM). The inverse modeling component leverages the trained GNS, reverse-mode automatic differentiation, and L-BFGS-B optimization to infer material properties from a target runout geometry. Its performance is demonstrated by back-calculating the material strengths that led to failure-induced runout in a dam system composed of multiple materials. Results are obtained within minutes and show good agreement with the target strength values. The framework introduced in this study provides an efficient approach for forward runout assessments and inverse strength back-calculation in realistic slope systems.

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.

Desk Editor's Note private letter to a colleague

This extends GNS to multi-layered slopes with cohesion and adds a differentiable inverse step for strength back-calculation, but the training coverage for the dam case looks limited.

read the letter

This paper takes the graph neural network simulator approach for granular flows and extends it to multi-layered slope systems that include both friction and cohesion. It adds a differentiable inverse component to back out material properties from a given runout shape. The forward modeling part is trained on column collapse tests and some multi-layered examples, then tested on slope runout. It matches the multi-material flow behavior from MPM while running much faster. The inverse side uses the trained network with reverse-mode differentiation and L-BFGS-B optimization to recover the strengths that produced a target dam failure geometry. They get the answers in minutes and the values line up with what was used to generate the target. That practical combination for both prediction and inference in realistic settings is what is new. Earlier GNS work stayed with uniform materials and simpler shapes, so this moves it closer to engineering use cases. The main concern is whether the training set is broad enough. Column collapse plus a handful of multi-layered runs may not cover the specific shear interactions and boundary effects in the dam example. If the GNS has systematic differences from the underlying MPM in those areas, the optimizer could settle on parameters that compensate for the mismatch instead of the actual material values. The abstract talks about good agreement but does not include specific error measures or tests on how the model generalizes beyond the training distributions. This kind of work is useful for geotechnical engineers and modelers who deal with slope stability and runout assessments. Readers who want faster surrogates or a way to estimate strengths from failure observations will see the most direct benefit. The extension is solid enough on its own terms to merit a full review, though the authors will need to strengthen the validation section. I would send this to peer review.

Referee Report

2 major / 1 minor

Summary. The manuscript introduces a differentiable Graph Neural Network Simulator (GNS) framework for forward and inverse modeling of multi-layered slope systems with multiple material properties. The forward component is a fine-tuned GNS incorporating friction angle and cohesion, trained on column collapse and selected multi-layered cases to replicate Material Point Method (MPM) runout dynamics with claimed computational speedup. The inverse component applies reverse-mode automatic differentiation and L-BFGS-B optimization on the trained GNS to back-calculate material strengths from target runout geometries, demonstrated on a dam system example with results obtained in minutes and reported good agreement with target values.

Significance. If the accuracy and generalization claims hold, the framework could enable rapid forward runout assessments and practical inverse strength back-calculation in complex multi-material geotechnical systems, where traditional MPM simulations are prohibitively slow. The explicit use of differentiability to support gradient-based optimization for inverse inference is a clear technical strength that distinguishes this from non-differentiable surrogate approaches.

major comments (2)

[Abstract] Abstract and results descriptions: the claims of 'significant computational speedup' and 'good agreement with the target strength values' are presented without any quantitative metrics (e.g., relative L2 errors, R² values, or wall-clock speedup factors with error bars). This absence is load-bearing because the central contribution is the framework's efficiency and reliability for both forward and inverse tasks.
[Inverse modeling component] Training and inverse demonstration sections: the GNS is fine-tuned primarily on column collapse plus a limited set of multi-layered cases. For the dam-system inverse example, there is no reported cross-validation, ablation on training diversity, or analysis of boundary/inter-layer shear effects specific to that geometry. Any systematic mismatch between GNS predictions and underlying MPM dynamics would cause L-BFGS-B to converge to compensating rather than physically correct parameters, directly affecting the inverse claim.

minor comments (1)

[Results] The manuscript would benefit from a dedicated table or figure panel directly comparing GNS vs. MPM runout profiles and material-property recovery errors across the demonstrated cases.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive feedback on our manuscript arXiv:2504.15938. We are pleased that the referee recognizes the significance of our differentiable GNS framework for forward and inverse modeling in multi-layered slope systems. We address the major comments below and commit to revisions that will enhance the clarity and robustness of our claims.

read point-by-point responses

Referee: [Abstract] Abstract and results descriptions: the claims of 'significant computational speedup' and 'good agreement with the target strength values' are presented without any quantitative metrics (e.g., relative L2 errors, R² values, or wall-clock speedup factors with error bars). This absence is load-bearing because the central contribution is the framework's efficiency and reliability for both forward and inverse tasks.

Authors: We agree with the referee that quantitative metrics are essential to substantiate the claims of computational speedup and agreement with target values. Although the manuscript describes the speedup qualitatively and notes that results are obtained within minutes, we will revise the abstract and results sections to include specific quantitative metrics. For the forward modeling, we will report the relative L2 error between GNS and MPM simulations, as well as wall-clock time comparisons with speedup factors. For the inverse modeling, we will provide the error metrics for the back-calculated material properties. These additions will be made in the revised version. revision: yes
Referee: [Inverse modeling component] Training and inverse demonstration sections: the GNS is fine-tuned primarily on column collapse plus a limited set of multi-layered cases. For the dam-system inverse example, there is no reported cross-validation, ablation on training diversity, or analysis of boundary/inter-layer shear effects specific to that geometry. Any systematic mismatch between GNS predictions and underlying MPM dynamics would cause L-BFGS-B to converge to compensating rather than physically correct parameters, directly affecting the inverse claim.

Authors: We appreciate the referee's concern regarding the training data diversity and potential for the optimizer to find compensating parameters in the inverse modeling. The GNS was trained on column collapse cases and selected multi-layered slope cases to incorporate the effects of multiple material properties including friction angle and cohesion. For the dam system demonstration, the inverse optimization yielded parameters in good agreement with targets, which we interpret as evidence that the GNS captures the relevant dynamics sufficiently for this purpose. However, to directly address the possibility of systematic mismatches, we will include in the revision an ablation study on the impact of training set diversity and a specific analysis of boundary and inter-layer shear effects for the dam geometry. This will strengthen the validation of the inverse component. revision: yes

Circularity Check

0 steps flagged

No significant circularity; forward training and inverse optimization are independent of target results

full rationale

The paper trains a GNS surrogate on MPM-generated data for column collapse and selected multi-layered cases to approximate forward runout dynamics, then applies the differentiable surrogate plus L-BFGS-B to recover material parameters from a held-out dam-system target geometry. Neither step reduces to its own inputs by construction: the forward model is fitted to external simulation data rather than self-defined, and the inverse step performs optimization to match an independent target rather than renaming a fitted quantity as a prediction. No self-citation load-bearing steps, uniqueness theorems, or ansatz smuggling appear in the provided abstract or description. The framework is therefore self-contained against external MPM benchmarks and receives a normal non-circularity finding.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The central claims rest on the assumption that a graph neural network can be trained to faithfully reproduce MPM physics for the chosen material models and geometries, plus standard assumptions of differentiability and convergence of the optimizer.

free parameters (1)

GNS network weights and hyperparameters
Learned during fine-tuning on simulation data; central to both forward accuracy and inverse results.

axioms (2)

domain assumption The underlying MPM simulations used for training provide ground-truth multi-material flow behavior.
Invoked when the GNS is positioned as a surrogate for MPM.
standard math Reverse-mode automatic differentiation through the GNS yields usable gradients for L-BFGS-B optimization.
Required for the inverse modeling component.

pith-pipeline@v0.9.0 · 5742 in / 1349 out tokens · 32363 ms · 2026-05-22T18:51:30.752498+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

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

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

GNS predicts the physical state... Yt = DΘ(Xt) ... Xt+1 = U(Xt, Yt) ... encoder–processor–decoder with MLPs... fine-tuned on column collapse and multi-layered slopes... L-BFGS-B optimization to infer material properties from target runout geometry
IndisputableMonolith/Foundation/AlphaCoordinateFixation.lean J_uniquely_calibrated_via_higher_derivative unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Mohr-Coulomb materials, w = [ϕ, c]... undrained shear strength back-calculation in dam system

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.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

Parameter-Efficient Conditioning for Material Generalization in Graph-Based Simulators
cs.LG 2025-11 unverdicted novelty 6.0

FiLM conditioning targeted at early message-passing layers lets pretrained GNS models generalize to new material properties using only 12 trajectories, a 5-fold data reduction versus multi-task baselines.