Extracting Interaction Kernels for Many-Particle Systems by a Two-Phase Approach

Liu Hong; Wuyue Yang; Yangxuan Shi

REVIEW 3 major objections 6 minor 17 references

Extracting Interaction Kernels for Many-Particle Systems by a Two-Phase Approach

T0 review · 3 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read This paper claims that unknown pairwise interaction rules in stochastic many-particle systems can be recovered directly from trajectory data by converting snapshots into a density and fitting the mean-field equation with a sparse…

desk verdict A plausible speedup-and-sparsity trick for mean-field kernel inversion, but the advertised power-law reconstruction is not supported as reported. read the letter →

arxiv 2501.00714 v1 pith:JST5OTQG submitted 2025-01-01 physics.comp-ph nlin.AOphysics.flu-dyn

Yangxuan Shi , Wuyue Yang , Liu Hong This is my paper · ORCID

classification physics.comp-phnlin.AOphysics.flu-dyn

keywords interactionkernelsmany-particlesystemsstochasticdynamicsmean-fieldequationsparseregressionimportancesamplingkerneldensityestimationmachinelearning

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper aims to establish that the interaction kernel of a stochastic many-particle system can be extracted from raw trajectory data without knowing the force law in advance. Its pipeline turns particle trajectories into a density via kernel density estimation, treats that density as a solution of the McKean–Vlasov mean-field equation, and solves a sparse regression problem for the kernel coefficients. A first phase uses importance-weighted time sampling and an adaptive threshold to select which dictionary terms matter, and a second phase refines those coefficients using the full dataset. If the method works as claimed, experimental or simulated trajectories alone would yield explicit force laws such as cubic potentials, singular repulsion–attraction forces, and piecewise opinion dynamics, with coefficients matching the true values to within a few percent.

What carries the argument

The load-bearing object is the constant-diffusion McKean–Vlasov mean-field equation $\partial_t u + \nabla_x\cdot[u(K*u)] = \frac{\sigma_c^2}{2}\Delta u$, which becomes linear in the coefficients once the kernel is written as $K(x,y)=\varphi(\|x-y\|_2)\frac{x-y}{\|x-y\|_2}$ and $\varphi$ is expanded in a dictionary. Kernel density estimation supplies the density $u$; finite differences supply $\partial_t u$ and $\Delta u$; discrete convolution supplies $K*u$. Phase I samples time steps with probability proportional to their current residual error, updates coefficients, and applies the threshold $\tau_k=\alpha\max_i|\zeta_i^{(k)}|$ to build a binary mask; Phase II fixes that mask and solves the full-data least-squares problem with an $\ell^1$ penalty.

What would settle it

Simulate a system with an interaction kernel that is not in the chosen dictionary, for example $\varphi(r)=r^{-1.5}$ using only integer-power basis functions, and monitor the mean-field residual after Phase II. The learned coefficient $-0.815$ instead of $-1$ shows that no coefficient vector can represent the true kernel; if the residual floor stays well above machine precision, the method is recovering the best dictionary fit rather than the true kernel.

Watch

Extended reading notes

Core claim

The central claim is that the residual of the mean-field equation, computed from kernel-density-estimated densities, is linear in the kernel coefficients, so identifying the interaction kernel becomes a sparse linear regression problem. Using a polynomial or physics-informed dictionary, Phase I selects active terms with importance sampling and adaptive thresholding, and Phase II refines the masked coefficients over all time steps. The reported results include a cubic kernel $\varphi(r)=3r^2$ recovered as $3.003$, a power-law repulsion–attraction kernel with coefficients $0.986$ and $-0.815$ (true values $1$ and $-1$), piecewise opinion kernels recovered to within about $0.01$, and a two-dimensional radial cubic kernel recovered as $2.916$ (true $3$). The same pipeline also reproduces free-energy curves, with Wasserstein distances between true and learned densities staying below $10^{-2}$ in the tested cases.

Load-bearing premise

Everything rests on the assumption that the density estimated from a finite number of particle snapshots is close enough to a smooth solution of the mean-field equation that its numerical derivatives and convolutions have small error, which requires many particles, a fine grid, and a small time step.

Editorial extensions

If this is right

Explicit interaction laws can be inferred from trajectory data alone, so downstream simulations can use the learned coefficients rather than a fitted black-box model.
Importance sampling reduces the Phase I computation time by about an order of magnitude, making large candidate dictionaries feasible.
The method handles smooth, singular, discontinuous, and two-dimensional radially symmetric kernels, so it applies to a broad class of pairwise interaction models.
Accuracy improves sharply with particle number, with Wasserstein error dropping by more than two orders of magnitude from $N=5000$ to $N=20000$, confirming the mean-field limit as the right bridge for large systems.
Macroscopic quantities such as free energy are also recoverable, allowing thermodynamic or collective-behavior comparisons beyond the kernel itself.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The method's accuracy is bounded by dictionary coverage: a kernel outside the chosen basis can only be approximated, as the power-law case's $-0.815$ coefficient shows when the true term is $r^{-1.5}$; moving to adaptive dictionaries or nonparametric bases is a natural next step.
The same residual-regression idea should transfer to weak-form or sample-based formulations for high-dimensional systems, a direction the authors themselves flag as future work.
If trajectories of interacting agents can be obtained, this pipeline offers a way to infer effective interaction rules for agent-based systems without knowing the rules a priori.
A testable extension would be to use the Phase I mask as a model-selection device across physically motivated dictionaries, such as Coulomb, Morse, or Lennard-Jones forms, on real experimental tracks.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Short version: the two-phase algorithm (importance-sampled adaptive thresholding, then frozen-mask refinement on all time steps) is a real, modest improvement over existing SINDy-style sparse regression for mean-field kernel extraction, and the clean recovery of the cubic and 2D radial cases is credible. But the power-law case is not demonstrated as claimed, and the missing code, hyperparameters, and baseline comparison make the paper hard to evaluate. It deserves a serious referee, not a desk reject, but it needs major revision.

What is new: the two-phase design is not in Lang & Lu (2022) or SINDy. Phase I using importance sampling on time steps with largest residual, plus an adaptive threshold to lock the sparsity mask, then Phase II refitting on all data with the mask frozen, is a sensible way to balance sparsity and accuracy. The discontinuity-location pre-step via DNN gradients is also a reasonable add-on for piecewise kernels. The numerical results for the cubic (3.003 vs 3) and 2D radial (2.916 vs 3) cases, and the piecewise linear opinion dynamics coefficients, are decent evidence the pipeline works when the kernel is in the dictionary.

Soft spots, in order of severity. First, the power-law validation is internally inconsistent. The dictionary is defined with integer exponents p, but the test kernel is r - r^{-1.5}. Table 1 lists a "True" coefficient of -1 in the x^{-p} column and learned -0.815. If p=1, the "True" entry is impossible; if p=1.5, the learned coefficient has an 18.5% error. Either way the singular-kernel claim is not supported. The paper should either use a dictionary that actually contains r^{-1.5}, or report the approximation error honestly. Second, hyperparameters (lambda, alpha, KDE bandwidth, the exponent p) are undisclosed, and no code or data is provided, so the experiments are not reproducible. Third, there is no baseline comparison against Lang & Lu 2022, the most directly relevant prior method, so the claimed speedup or accuracy gain is not quantified against the state of the art. Minor: Eq. (10) has a time-index slip (u(x,t) outside the time integral should be u(x,s)), and the Kac reference is missing in Theorem 2.3. These are fixable.

Who it is for: people working on data-driven discovery of interaction rules from particle trajectories or mean-field PDEs. The central idea is useful enough to warrant referee time, but the paper as written overclaims "accurate and robust" reconstruction across singular kernels. I would recommend sending to peer review with a request for major revision: fix the power-law experiment, add code/data and hyperparameters, and run a baseline comparison.

Referee Report

3 major / 6 minor

Summary. The paper proposes a two-phase regression method for extracting interaction kernels of stochastic many-particle systems from full particle trajectories. The trajectories are converted into a density field by kernel density estimation, and the residual of the mean-field PDE (Eq. 13) is minimized in Phase I with importance sampling and an adaptive threshold sparsification, after which Phase II refines the non-masked coefficients using all time steps. The method is demonstrated on one-dimensional systems with cubic, power-law, and piecewise-discontinuous kernels and on a two-dimensional radially symmetric cubic kernel, with comparisons based on learned coefficients, Wasserstein distance, free energy, and convergence plots.

Significance. If the results are valid, the proposal is a useful addition to sparse-identification methods for interaction kernels: it works directly from raw trajectories, uses a principled mean-field residual, and includes a sensible sparsification/refinement split. The clean recovery of the cubic coefficient (3.003 vs 3) in the one-dimensional case, the extension to a two-dimensional radially symmetric potential, and the system-level comparisons via Wasserstein distance and free energy are strong points, as is the explicit discussion of the method's limitations in high dimensions and for small data sets. The main reservations concern the consistency of the power-law validation in Table 1, missing hyperparameter and bandwidth disclosure, and a flaw in the displayed weak-form derivation; none of these is necessarily fatal, but they must be fixed before the advertised claims are fully supported.

major comments (3)

[Section 2, Eq. (10)] The displayed weak formulation is not correct. The time integral must act on the whole expression: the correct term is ∫_0^t ∫ g'(x) · [∫ K(x,y)u(y,s) dy] u(x,s) dx ds, followed by the analogous diffusion term. As written, Eq. (10) has u(x,t) outside the ds integral, so the subsequent step of differentiating with respect to t is not valid. Since Eq. (13) is the basis of the regression loss in Eq. (17), this derivation should be corrected.
[Section 4.2 and Table 1] The headline 'power-law repulsion-attraction' experiment is internally inconsistent. The true kernel is r - r^{-1.5}, but the dictionary D_poly in Section 3.1.1 contains only integer powers; moreover, the modified potential with cutoff radius r_c is constant on r ≤ r_c, so the exact kernel is not in the dictionary. Table 1 lists a 'True' coefficient of -1 for the x^{-p} term without stating p: if p=1 the target is not r^{-1.5}, and if p=1.5 the term is outside the declared dictionary. The learned coefficient -0.815 therefore does not demonstrate accurate reconstruction of the singular power-law kernel. Please either extend the dictionary to include p=1.5, rerun the experiment and report the fitted coefficient with its error, or explicitly analyze the projection error of the true kernel onto the dictionary and temper the corresponding claim in Section 5.
[Sections 3 and 4] The reproducibility of the method is blocked by missing numerical details. The KDE bandwidth, the finite-difference grid parameters, the regularization parameter λ, the sparsity intensity α, the optimization algorithm and stopping criterion for Phases I and II, and the exact dictionary used for the power-law case are not reported. Since the loss in Eq. (17) uses finite differences of the KDE density, the results can depend strongly on these choices; Figure 3 shows order-of-magnitude variation in accuracy with N and Δt. A complete table of hyperparameters for all four experiments should be provided.

minor comments (6)

[Section 5] The text 'power-low repulsion-attraction potential' contains a typo and should read 'power-law'.
[Section 2, Theorem 2.3] The citation 'Kac[?]' is an unresolved placeholder and should be completed.
[Table 1] The row labeled 'Case D' in Table 1 is not linked to any 'Case D' in Section 4.4; please harmonize the case labels between the table and the text.
[Section 3.1.2, Eq. (18)] The per-time-step error E_n used for importance sampling is never defined; please state its exact formula.
[Section 4.3] The DNN baseline is described only in a few sentences; please provide architecture, training details, or a reference so that the comparison is meaningful.
[General] The paper does not include a data/code availability statement; for a purely numerical methodology, releasing code and parameter settings would substantially aid reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the learned kernels are free parameters fitted to a PDE residual, with ground truth supplied externally.

full rationale

The paper's pipeline is a standard inverse-problem regression: it derives the mean-field PDE (Eq. 13) from the SDE dynamics, expands the unknown kernel in a dictionary (Eq. 16), and minimizes the PDE residual (Eqs. 17 and 21). The coefficients are free parameters optimized against data; they are not defined in terms of the target kernel values or normalized to force the reported results. Ground-truth coefficients (e.g., 3 for the cubic potential, 0.5/1.0 for the opinion kernels) are chosen independently and compared after fitting, so the comparisons are external rather than built into the loss. The method's use of the same mean-field equation family for both data generation and regression is a synthetic-validation convention, not a logical reduction: the equation does not encode the unknown coefficients. No load-bearing self-citation appears; references [12,13] are standard external results on McKean-Vlasov equations and propagation of chaos. The apparent inconsistency in Case B (dictionary with integer powers vs. a true r^-1.5 term, with Table 1 reporting -1/-0.815) is a validation/specification issue, not circularity.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

No new physical entities, forces, or dimensions are introduced; the two-phase algorithm is a procedure, not an entity. The free parameters are hyperparameters and modeling choices that are not disclosed, and the axioms are standard or domain-specific assumptions inherited from the mean-field framework.

free parameters (4)

alpha (sparsity intensity)
Eq. (19): threshold tau_k = alpha * max|zeta|; value not reported; controls which terms survive Phase I and therefore the final mask.
lambda (L1 regularization weight)
Eqs. (17) and (21): balances PDE residual against sparsity; value not reported.
KDE bandwidth
Required to estimate u(x,t) and its derivatives from particle positions in Section 3.1.1; never specified, though derivative accuracy controls the regression residual.
Exponent p in the x^{-p} dictionary term
Table 1 uses a single x^{-p} basis for the power-law case; the true term is r^{-1.5}, so p must be chosen near 1.5 for the coefficient 1.0 to be meaningful, but the chosen p is not stated.

assumptions (5)

domain assumption Propagation of chaos: as N -> infinity, the empirical measure converges to u(x,t) solving the McKean-Vlasov equation (Theorem 2.3).
Invoked in Section 2 to justify using Eq. (13) as the model for the KDE density; the paper cites 'Kac[?]' without a reference and acknowledges the derivation is formal.
domain assumption The interaction is pairwise radial with K(x,y) = grad Phi(||x-y||) and the diffusion is constant, so Eq. (13) holds.
Section 3 restricts to this form; the entire method fails for non-radial or state-dependent diffusion kernels.
ad hoc to paper The true kernel lies in the span of the chosen dictionary {psi_k}.
Section 3.1.1 expands phi(r) as sum zeta_k psi_k; if the true kernel is outside the dictionary, the regression cannot recover it; no identifiability or consistency analysis is given.
standard math Finite-difference approximations of derivatives and discrete convolution on the KDE grid accurately represent the continuous PDE residual.
Section 3.1.1 replaces partial derivatives and convolution by finite differences; no error bounds or mesh-convergence study is reported.
domain assumption The empirical density estimated by KDE is a sufficiently accurate solution of Eq. (13).
The method uses the KDE density in place of u in the PDE residual; the paper's Conclusion notes a large number of particles is required, and Fig. 3 shows strong dependence of accuracy on N and dt.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Extracting Interaction Kernels for Many-Particle Systems by a Two-Phase Approach." pith.science (2026). https://pith.science/paper/JST5OTQG

@misc{pith2026250100714,
  author       = {Pith},
  title        = {Pith review of: Extracting Interaction Kernels for Many-Particle Systems by a Two-Phase Approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JST5OTQG}},
  note         = {Machine review of arXiv:2501.00714}
}

read the original abstract

This paper presents a two-phase method for learning interaction kernels of stochastic many-particle systems. After transforming stochastic trajectories of every particle into the particle density function by the kernel density estimation method, the first phase of our approach combines importance sampling with an adaptive threshold strategy to identify key terms in the kernel function, while the second phase uses the whole dataset to refine the coefficients. During the implementation of our method, the mean-field equation plays a key role in reformulating the task of extracting the interaction kernels into a learnable regression problem. We demonstrate the outstanding performance of our approach through extensive numerical examples, including interacting particle systems with a cubic potential, power-law repulsion-attraction potential, piecewise linear potential, as well as a two-dimensional radially symmetric potential.

Figures

Figures reproduced from arXiv: 2501.00714 by the authors.

**Figure 2.** Identification of interaction kernels using the two-phase method in the case of cubic potential. Evolution of particle density from t = 0 to t = 1, calculated by (a) coupled stochastic differential equations and (b) the mean-field equation, respectively. (c) The learned interaction kernel φ(r) from either stochastic trajectories or the particle density with an inset showing their absolute errors. (d) Comparison on t… view at source ↗

**Figure 3.** Comparison on the predicted particle density uˆ(x,t) and the exact one u(x,t) based on their Wasserstein distance W2(u,uˆ). (a) The total number of interacting particles N, (b) time step size ∆t, (c) variance s 2 , and (d) diffusion coefficient σ 2 c are varied respectively. Through comparing Fig.2(a) and Fig.2(b), we can see an excellent agreement on the results of stochastic particle simulations and the mean-fiel… view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Results for the power-law repulsion-attraction potential. Evolution of particle density calculated by the (a) SDE and (b) the mean-field equation. (c) Comparison between true and learned interaction kernels φ(r). (d) Wasserstein distance W2(u,uˆ) between true and lear…

**Figure 5.** Figure 5: Opinion dynamics with discontinuities. Evolution of opinion density distribution showing (a) polarization and (e) consensus. (b,f) Comparison between true and learned interaction functions by using DNN. (c,g) Gradient of learned interaction functions f(r). (d,h) Compar…

**Figure 6.** Figure 6: 2D radial symmetric potential. (a) Evolution of the density distribution estimated based on stochastic simulations of 20,000 particles. (b) Free energy under different initial conditions. (c) Comparison between true and learned coefficients in the dictionary representa…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

17 extracted references · 13 canonical work pages

[1]

The rise and potential of large language model based agents: A survey

Zhiheng Xi, Wenxiang Chen, Xin Guo, Wei He, Yiwen Ding, Boyang Hong, Ming Zhang, Junzhe Wang, Senjie Jin, Enyu Zhou, et al. The rise and potential of large language model based agents: A survey. arXiv preprint arXiv:2309.07864, 2023

arXiv 2023
[2]

Discovering governing equations from data by sparse identification of nonlinear dynamical systems.Proceedings of the national academy of sciences, 113(15):3932–3937, 2016

Steven L Brunton, Joshua L Proctor, and J Nathan Kutz. Discovering governing equations from data by sparse identification of nonlinear dynamical systems.Proceedings of the national academy of sciences, 113(15):3932–3937, 2016

2016
[3]

Learning anisotropic interaction rules from individual trajectories in a heterogeneous cellular population

Daniel A Messenger, Graycen E Wheeler, Xuedong Liu, and David M Bortz. Learning anisotropic interaction rules from individual trajectories in a heterogeneous cellular population. Journal of the Royal Society Interface, 19(195):20220412, 2022

work page 2022
[4]

Learning hydrodynamic equations for active matter from particle simulations and experiments

Rohit Supekar, Boya Song, Alasdair Hastewell, Gary PT Choi, Alexander Mietke, and Jörn Dunkel. Learning hydrodynamic equations for active matter from particle simulations and experiments. Pro- ceedings of the National Academy of Sciences, 120(7):e2206994120, 2023

work page 2023
[5]

Learning interaction kernels in mean-field equations of first-order systems of interacting particles

Quanjun Lang and Fei Lu. Learning interaction kernels in mean-field equations of first-order systems of interacting particles. SIAM Journal on Scientific Computing, 44(1):A260–A285, 2022. 17

work page 2022
[6]

Multiscale modeling meets machine learning: What can we learn? Archives of Computational Methods in Engi- neering, 28:1017–1037, 2021

Grace CY Peng, Mark Alber, Adrian Buganza Tepole, William R Cannon, Suvranu De, Savador Dura- Bernal, Krishna Garikipati, George Karniadakis, William W Lytton, Paris Perdikaris, et al. Multiscale modeling meets machine learning: What can we learn? Archives of Computational Methods in Engi- neering, 28:1017–1037, 2021

work page 2021
[7]

Multiscale simulations of complex systems by learning their effective dynamics

Pantelis R Vlachas, Georgios Arampatzis, Caroline Uhler, and Petros Koumoutsakos. Multiscale simulations of complex systems by learning their effective dynamics. Nature Machine Intelligence, 4(4):359–366, 2022

2022
[8]

Physi- cally informed data-driven modeling of active nematics

Matthew Golden, Roman O Grigoriev, Jyothishraj Nambisan, and Alberto Fernandez-Nieves. Physi- cally informed data-driven modeling of active nematics. Science Advances, 9(27):eabq6120, 2023

work page 2023

Show all 17 references

[9]

Mod- eling collective motion: variations on the vicsek model.The European Physical Journal B, 64:451–456, 2008

Hugues Chaté, Francesco Ginelli, Guillaume Grégoire, Fernando Peruani, and Franck Raynaud. Mod- eling collective motion: variations on the vicsek model.The European Physical Journal B, 64:451–456, 2008

2008
[10]

Traffic and related self-driven many-particle systems

Dirk Helbing. Traffic and related self-driven many-particle systems. Reviews of modern physics , 73(4):1067, 2001

2001
[11]

Heterophilious dynamics enhances consensus

Sebastien Motsch and Eitan Tadmor. Heterophilious dynamics enhances consensus. SIAM review, 56(4):577–621, 2014

2014
[12]

A class of markov processes associated with nonlinear parabolic equations

Henry P McKean Jr. A class of markov processes associated with nonlinear parabolic equations. Proceedings of the National Academy of Sciences, 56(6):1907–1911, 1966

1907
[13]

Topics in propagation of chaos

Alain-Sol Sznitman. Topics in propagation of chaos. Ecole d’été de probabilités de Saint-Flour XIX—1989, 1464:165–251, 1991

1989
[14]

Maxwell’s equations

Paul G Huray. Maxwell’s equations. John Wiley & Sons, 2009

2009
[15]

Solving inverse stochastic prob- lems from discrete particle observations using the fokker–planck equation and physics-informed neural networks

Xiaoli Chen, Liu Yang, Jinqiao Duan, and George Em Karniadakis. Solving inverse stochastic prob- lems from discrete particle observations using the fokker–planck equation and physics-informed neural networks. SIAM Journal on Scientific Computing, 43(3):B811–B830, 2021

2021
[16]

Weak collocation regression method: Fast reveal hidden stochastic dynamics from high-dimensional aggregate data

Liwei Lu, Zhijun Zeng, Yan Jiang, Yi Zhu, and Pipi Hu. Weak collocation regression method: Fast reveal hidden stochastic dynamics from high-dimensional aggregate data. Journal of Computational Physics, 502:112799, 2024

2024
[17]

Automating the search for artificial life with foundation models

Akarsh Kumar, Chris Lu, Louis Kirsch, Yujin Tang, Kenneth O Stanley, Phillip Isola, and David Ha. Automating the search for artificial life with foundation models. arXiv preprint arXiv:2412.17799 , 2024. 18

2024 arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [1]

The rise and potential of large language model based agents: A survey

Zhiheng Xi, Wenxiang Chen, Xin Guo, Wei He, Yiwen Ding, Boyang Hong, Ming Zhang, Junzhe Wang, Senjie Jin, Enyu Zhou, et al. The rise and potential of large language model based agents: A survey. arXiv preprint arXiv:2309.07864, 2023

arXiv 2023

[2] [2]

Discovering governing equations from data by sparse identification of nonlinear dynamical systems.Proceedings of the national academy of sciences, 113(15):3932–3937, 2016

Steven L Brunton, Joshua L Proctor, and J Nathan Kutz. Discovering governing equations from data by sparse identification of nonlinear dynamical systems.Proceedings of the national academy of sciences, 113(15):3932–3937, 2016

2016

[3] [3]

Learning anisotropic interaction rules from individual trajectories in a heterogeneous cellular population

Daniel A Messenger, Graycen E Wheeler, Xuedong Liu, and David M Bortz. Learning anisotropic interaction rules from individual trajectories in a heterogeneous cellular population. Journal of the Royal Society Interface, 19(195):20220412, 2022

work page 2022

[4] [4]

Learning hydrodynamic equations for active matter from particle simulations and experiments

Rohit Supekar, Boya Song, Alasdair Hastewell, Gary PT Choi, Alexander Mietke, and Jörn Dunkel. Learning hydrodynamic equations for active matter from particle simulations and experiments. Pro- ceedings of the National Academy of Sciences, 120(7):e2206994120, 2023

work page 2023

[5] [5]

Learning interaction kernels in mean-field equations of first-order systems of interacting particles

Quanjun Lang and Fei Lu. Learning interaction kernels in mean-field equations of first-order systems of interacting particles. SIAM Journal on Scientific Computing, 44(1):A260–A285, 2022. 17

work page 2022

[6] [6]

Multiscale modeling meets machine learning: What can we learn? Archives of Computational Methods in Engi- neering, 28:1017–1037, 2021

Grace CY Peng, Mark Alber, Adrian Buganza Tepole, William R Cannon, Suvranu De, Savador Dura- Bernal, Krishna Garikipati, George Karniadakis, William W Lytton, Paris Perdikaris, et al. Multiscale modeling meets machine learning: What can we learn? Archives of Computational Methods in Engi- neering, 28:1017–1037, 2021

work page 2021

[7] [7]

Multiscale simulations of complex systems by learning their effective dynamics

Pantelis R Vlachas, Georgios Arampatzis, Caroline Uhler, and Petros Koumoutsakos. Multiscale simulations of complex systems by learning their effective dynamics. Nature Machine Intelligence, 4(4):359–366, 2022

2022

[8] [8]

Physi- cally informed data-driven modeling of active nematics

Matthew Golden, Roman O Grigoriev, Jyothishraj Nambisan, and Alberto Fernandez-Nieves. Physi- cally informed data-driven modeling of active nematics. Science Advances, 9(27):eabq6120, 2023

work page 2023