Meta-Inverse Physics-Informed Neural Networks for High-Dimensional Ordinary Differential Equations

Chin Chun Ooi; James Chun Yip Chan; Kenneth Hor Cheng Koh; Sheng Yuan Chin; Yew-Soon Ong; Zhao Wei

arxiv: 2605.03511 · v1 · submitted 2026-05-05 · 💻 cs.LG · cs.AI

Meta-Inverse Physics-Informed Neural Networks for High-Dimensional Ordinary Differential Equations

Zhao Wei , Kenneth Hor Cheng Koh , Sheng Yuan Chin , James Chun Yip Chan , Chin Chun Ooi , Yew-Soon Ong This is my paper

Pith reviewed 2026-05-07 17:11 UTC · model grok-4.3

classification 💻 cs.LG cs.AI

keywords meta-learningphysics-informed neural networksinverse problemsordinary differential equationspharmacokinetic modelsparameter estimationhigh-dimensional dynamical systemsmulti-scale dynamics

0 comments

The pith

MI-PINN recovers unknown kinetic parameters and missing terms in systems of up to 33 coupled ODEs by first learning a shared physics representation across tasks then fixing it for each new inverse problem.

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

The paper presents MI-PINN to address inverse problems in high-dimensional coupled ODEs where data are sparse and some dynamics or parameters are unknown. It casts the task as two-stage meta-learning: a neural network first learns a physics-aware representation from multiple related tasks, then for any new task only the task-specific unknowns are optimized while the representation stays fixed. This lowers the dimension of the search space and improves accuracy under limited observations. An adaptive clustering multi-branch scheme is added to capture multi-scale dynamics. The approach is shown on whole-body PBPK models for paracetamol and theophylline under different dosing routes, recovering masked parameters and reconstructing absent mechanistic terms.

Core claim

MI-PINN reformulates inverse modeling for high-dimensional ODE systems as a two-stage meta-learning problem. A physics-aware neural representation is learned across multiple tasks; for each new inverse task the representation is held fixed while only task-specific unknowns are optimized. Combined with an adaptive clustering-based multi-branch scheme for multi-scale dynamics, this enables accurate recovery of masked kinetic parameters and reconstruction of missing mechanistic terms from limited clinical observations in PBPK models containing up to 33 coupled ODEs.

What carries the argument

Two-stage meta-learning in which a shared physics-aware representation is learned across tasks and then fixed while optimizing only the task-specific unknowns for each inverse problem, thereby reducing the effective parameter dimension.

If this is right

Accurate recovery of kinetic parameters becomes possible even when clinical observations cover only a few measurable channels.
Missing mechanistic terms in the ODE system can be reconstructed without jointly optimizing the entire model from scratch.
Sample efficiency improves because the search dimension per task is reduced by keeping the learned representation fixed.
Multi-scale dynamics in high-dimensional ODEs are handled through adaptive clustering into separate network branches.
The same fixed representation supports inference across different dosing scenarios and compounds within the PBPK family.

Where Pith is reading between the lines

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

The method could support rapid personalization of pharmacokinetic models using only a handful of measurements from an individual patient.
If the learned representation captures sufficiently general physical structure, it might transfer to inverse problems in entirely different dynamical domains such as chemical kinetics or ecological models.
Real-time monitoring devices could update patient-specific parameters by solving only a low-dimensional optimization while reusing the pre-learned representation.
Model development for new drugs could require fewer dedicated experiments once a library of representations from prior compounds exists.

Load-bearing premise

A physics-aware representation learned from multiple tasks will stay effective and generalizable when held fixed for new inverse tasks that have only sparse, partially observed data in high-dimensional ODE systems.

What would settle it

On a new PBPK model with known ground-truth parameters, fixing the learned representation and optimizing only the unknowns yields large errors in recovered parameter values or fails to reconstruct deliberately omitted mechanistic terms.

read the original abstract

Solving inverse problems in dynamical systems governed by high-dimensional coupled ordinary differential equations (ODEs) is a ubiquitous challenge in scientific machine learning. In many real-world applications, researchers seek to uncover unknown parameters or model unknown dynamics even as the underlying physics is only partially characterized, and observations are sparse and limited to specific measurable channels. While physics-informed neural networks (PINNs) are ideal for inverse inference under partial observability, existing PINNs typically rely on task-specific joint optimization, which suffers from optimization difficulties and poor generalization. In this paper, we propose a meta-inverse physics-informed neural network (MI-PINN) that reformulates inverse modeling as a two-stage meta-learning problem. MI-PINN first learns a physics-aware representation across multiple tasks, and then performs inverse modeling by optimizing task-specific unknowns while keeping the learned representation fixed. This two-stage formulation significantly reduces the parameter search dimension, thereby improving sample efficiency and enabling accurate inference. To handle multi-scale dynamics common in these high-dimensional ODE systems, we further introduce an adaptive clustering-based multi-branch learning scheme. We demonstrate the effectiveness of MI-PINN on whole-body physiologically based pharmacokinetic (PBPK) models with up to 33 coupled ODEs, using paracetamol and theophylline under intravenous and oral dosing scenarios. Experimental results show that MI-PINN enables accurate recovery of masked kinetic parameters and reconstruction of missing mechanistic terms despite limited clinical observations.

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

MI-PINN uses a two-stage meta-learning setup to freeze a shared network after training and optimize only unknowns for inverse problems in high-dimensional ODEs, which helps on the PBPK examples but depends on the meta-features covering new dynamics.

read the letter

The main takeaway is that this paper reframes inverse PINNs as a meta-learning task: first train a physics-aware network across multiple related problems, then freeze its weights and solve only for the task-specific parameters or missing terms. That reduces the optimization load for systems with dozens of coupled ODEs and sparse observations, which matches the PBPK use case they target. The adaptive clustering multi-branch part is a practical addition for handling different time scales within those models. They show it recovering masked kinetic parameters and filling in missing mechanistic terms on paracetamol and theophylline models up to 33 ODEs under IV and oral dosing with limited clinical data. That is concrete and better than pure theory or low-dimensional toys. The experiments appear to support the claim on the reported instances. The soft spot is the core assumption that the fixed meta-representation stays accurate enough when applied to new tasks whose parameters or structure differ from the meta-training distribution. If a new PBPK instance pushes trajectories outside what the frozen network can approximate, the physics loss will not drive reliable recovery of the unknowns. The stress-test note flags exactly this, and the abstract gives no ablations on how far the test cases deviate or how sensitive results are to the meta-training set. I would want to see those checks in the full text. This is aimed at researchers doing inverse modeling in systems biology or pharmacokinetics who already know PINNs and want a way to scale them without joint optimization on every new problem. A reader working on meta-learning for scientific ML would find the formulation and the clustering scheme useful. It deserves a serious referee because the problem is real, the method is a clear step beyond standard inverse PINNs, and the application is grounded. I would send it to review rather than desk reject, expecting questions mainly on generalization and baseline comparisons.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes Meta-Inverse Physics-Informed Neural Networks (MI-PINN) to solve inverse problems for high-dimensional coupled ODE systems, such as whole-body PBPK models with up to 33 equations. It reformulates inverse modeling as a two-stage meta-learning task: a shared physics-aware neural representation is first learned across multiple tasks, after which task-specific unknowns (masked kinetic parameters or missing mechanistic terms) are optimized while the representation weights are held fixed. An adaptive clustering multi-branch scheme is introduced to address multi-scale dynamics. Experiments on paracetamol and theophylline PBPK models under intravenous and oral dosing demonstrate recovery of masked parameters and reconstruction of missing terms from limited, partially observed clinical data.

Significance. If the central claims hold, MI-PINN would offer a practical advance for inverse inference in sparse-data, high-dimensional dynamical systems by reducing the effective optimization dimension through meta-learned representations. The two-stage separation and the multi-branch clustering for multi-scale PBPK dynamics address real limitations of standard PINNs. The application to clinically relevant models with up to 33 coupled ODEs provides a concrete test bed; reproducible code or parameter-free derivations are not mentioned, but the empirical focus on recovery under partial observability is a strength if properly benchmarked.

major comments (2)

[§3] §3: The procedure explicitly optimizes only the task-specific unknowns while freezing the meta-trained network weights. This load-bearing step assumes the learned representation spans the solution manifold for new tasks whose parameters or missing terms may shift trajectories outside the meta-training distribution. No theoretical characterization or out-of-distribution ablation is provided to bound the approximation error for 33-ODE systems, leaving the reliability of the subsequent physics-loss minimization unverified.
[§4] §4 (experimental results): The reported recovery of masked kinetic parameters and missing terms on the paracetamol/theophylline PBPK models lacks explicit baselines (standard PINN or meta-PINN variants), quantitative error metrics with confidence intervals, statistical significance tests, and component ablations (meta-learning stage vs. clustering). Without these, the claimed gains in sample efficiency and accuracy cannot be assessed as load-bearing evidence for the central claim.

minor comments (2)

[§3] Notation for the adaptive clustering scheme and the two-stage loss could be clarified with a single consolidated diagram or pseudocode to aid reproducibility.
[Abstract] The abstract states 'accurate recovery' without defining the quantitative threshold; a brief statement of the error tolerance used in the experiments would improve precision.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment point by point below and outline the revisions we will make.

read point-by-point responses

Referee: [§3] The procedure explicitly optimizes only the task-specific unknowns while freezing the meta-trained network weights. This load-bearing step assumes the learned representation spans the solution manifold for new tasks whose parameters or missing terms may shift trajectories outside the meta-training distribution. No theoretical characterization or out-of-distribution ablation is provided to bound the approximation error for 33-ODE systems, leaving the reliability of the subsequent physics-loss minimization unverified.

Authors: We agree that a formal theoretical characterization of the approximation error would strengthen the claims, but deriving tight bounds for 33-dimensional nonlinear ODE systems remains an open challenge beyond the scope of this work. Our method is motivated by empirical generalization observed in the PBPK experiments. To address the concern directly, we will add an out-of-distribution ablation study in the revised manuscript that systematically varies kinetic parameters and missing terms outside the meta-training distribution and reports the resulting physics-loss and recovery errors. This will provide concrete empirical evidence on the reliability of the frozen representation. revision: partial
Referee: [§4] The reported recovery of masked kinetic parameters and missing terms on the paracetamol/theophylline PBPK models lacks explicit baselines (standard PINN or meta-PINN variants), quantitative error metrics with confidence intervals, statistical significance tests, and component ablations (meta-learning stage vs. clustering). Without these, the claimed gains in sample efficiency and accuracy cannot be assessed as load-bearing evidence for the central claim.

Authors: We concur that the current experimental section would benefit from more rigorous controls and statistical reporting. In the revised manuscript we will include direct comparisons to standard PINN and meta-PINN baselines, report mean absolute errors with 95% confidence intervals computed over multiple random seeds, apply appropriate statistical significance tests (e.g., paired t-tests), and add component ablations that isolate the meta-learning stage from the adaptive clustering scheme. These additions will allow readers to quantitatively evaluate the claimed improvements in sample efficiency and accuracy. revision: yes

Circularity Check

0 steps flagged

No significant circularity; method is a self-contained reformulation

full rationale

The paper reformulates inverse ODE problems as a two-stage meta-learning procedure: a shared physics-aware network is trained across tasks, then frozen while task-specific unknowns (parameters or missing terms) are optimized. This is presented as an algorithmic design choice that reduces search dimensionality, with an additional adaptive clustering multi-branch scheme for multi-scale dynamics. No derivation chain equates outputs to inputs by construction, no predictions are statistically forced from fitted subsets, and no load-bearing uniqueness theorems or ansatzes are imported via self-citation. Effectiveness is asserted via experiments on PBPK models (up to 33 ODEs) rather than tautological definitions. The approach is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 1 invented entities

Based solely on the abstract, the paper relies on standard assumptions of PINNs (neural networks can approximate ODE solutions) and meta-learning (shared representations transfer across tasks); no explicit free parameters or invented physical entities are detailed.

axioms (2)

standard math Neural networks can approximate solutions and inverses of ODE systems under partial observability
Core premise of physics-informed neural networks referenced in the abstract
domain assumption A representation learned across multiple tasks captures generalizable physics for new tasks
Central to the two-stage meta-learning formulation described

invented entities (1)

MI-PINN architecture no independent evidence
purpose: Two-stage meta-inverse learning framework for ODE inverse problems
New method proposed in the paper

pith-pipeline@v0.9.0 · 5566 in / 1407 out tokens · 82413 ms · 2026-05-07T17:11:47.129118+00:00 · methodology

Review history (2 revisions) →

discussion (0)

Reference graph

Works this paper leans on

9 extracted references · 8 canonical work pages

[1]

A repository of protein abundance data of drug metabolizing enzymes and transporters for applications in physiologically based pharmacokinetic (PBPK) modelling and simulation,

M. K. Ladumor, A. Thakur, S. Sharma et al., “A repository of protein abundance data of drug metabolizing enzymes and transporters for applications in physiologically based pharmacokinetic (PBPK) modelling and simulation,” Sci. Rep., vol. 9, no. 1, Art. no. 9709, Jul. 2019, doi: 10.1038/s41598-019-46075-8

work page doi:10.1038/s41598-019-46075-8 2019
[2]

Prediction of human volume of distribution values for neutral and basic drugs. 2. Extended data set and leave-class- out statistics,

F. Lombardo, R. S. Obach, M. Y . Shalaeva, and F. Gao, “Prediction of human volume of distribution values for neutral and basic drugs. 2. Extended data set and leave-class- out statistics,” J. Med. Chem. , vol. 47, no. 5, pp. 1242 –1250, Feb. 2004. doi: 10.1021/jm030408h

work page doi:10.1021/jm030408h 2004
[3]

Evaluation of whole blood theophylline enzyme immunochromatography assay,

M. P. Habib, R. B. Schifman, B. Y . Shon, J. F. Fiastro, and S. C. Campbell, “Evaluation of whole blood theophylline enzyme immunochromatography assay,” Chest, vol. 92, no. 1, pp. 129–131, Jul. 1987. doi: 10.1378/chest.92.1.129

work page doi:10.1378/chest.92.1.129 1987
[4]

Hansch, A

C. Hansch, A. Leo, and D. Hoekman, Exploring QSAR: Hydrophobic, Electronic, and Steric Constants, vol. 2. Washington, DC, USA: Amer. Chem. Soc., 1995

1995
[5]

Caco-2 cell line standardization with pharmaceutical requirements and in vitro model suitability for permeability assays,

M. Kus, I. Ibragimow, and H. Piotrowska-Kempisty, “Caco-2 cell line standardization with pharmaceutical requirements and in vitro model suitability for permeability assays,” Pharmaceutics, vol. 15, no. 11, Art. no. 2523, Oct. 2023. doi: 10.3390/pharmaceutics15112523

work page doi:10.3390/pharmaceutics15112523 2023
[6]

Optimization of physicochemical properties of theophylline by forming cocrystals with amino acids,

Q. Li, Y . Xie, Z. Wang et al. , “Optimization of physicochemical properties of theophylline by forming cocrystals with amino acids,” RSC Adv., vol. 14, no. 54, pp. 40006–40017, Nov. 2024. doi: 10.1039/d4ra06804a

work page doi:10.1039/d4ra06804a 2024
[7]

Clinical pharmacodynamics of theophylline,

C. C. Peck, A. I. Nichols, J. Baker, L. L. Lenert, and D. Ezra, “Clinical pharmacodynamics of theophylline,” J. Allergy Clin. Immunol., vol. 76, no. 2, pp. 292– 297, Aug. 1985. doi: 10.1016/0091-6749(85)90644-x

work page doi:10.1016/0091-6749(85)90644-x 1985
[8]

Metabolism of theophylline by cDNA-expressed human cytochromes P -450,

H. R. Ha, J. Chen, A. U. Freiburghaus, and F. Follath, “Metabolism of theophylline by cDNA-expressed human cytochromes P -450,” Br. J. Clin. Pharmacol., vol. 39, no. 3, pp. 321–326, Mar. 1995. doi: 10.1111/j.1365-2125.1995.tb04455.x

work page doi:10.1111/j.1365-2125.1995.tb04455.x 1995
[9]

Abduljalil, I

K. Abduljalil, I. Gardner, and M. Jamei, “Application of a physiologically based pharmacokinetic approach to predict theophylline pharmacokinetics using virtual non- pregnant, pregnant, fetal, breast-feeding, and neonatal populations,” Front. Pediatr., vol. 10, Art. no. 840710, Mar. 2022. doi: 10.3389/fped.2022.840710

work page doi:10.3389/fped.2022.840710 2022

[1] [1]

A repository of protein abundance data of drug metabolizing enzymes and transporters for applications in physiologically based pharmacokinetic (PBPK) modelling and simulation,

M. K. Ladumor, A. Thakur, S. Sharma et al., “A repository of protein abundance data of drug metabolizing enzymes and transporters for applications in physiologically based pharmacokinetic (PBPK) modelling and simulation,” Sci. Rep., vol. 9, no. 1, Art. no. 9709, Jul. 2019, doi: 10.1038/s41598-019-46075-8

work page doi:10.1038/s41598-019-46075-8 2019

[2] [2]

Prediction of human volume of distribution values for neutral and basic drugs. 2. Extended data set and leave-class- out statistics,

F. Lombardo, R. S. Obach, M. Y . Shalaeva, and F. Gao, “Prediction of human volume of distribution values for neutral and basic drugs. 2. Extended data set and leave-class- out statistics,” J. Med. Chem. , vol. 47, no. 5, pp. 1242 –1250, Feb. 2004. doi: 10.1021/jm030408h

work page doi:10.1021/jm030408h 2004

[3] [3]

Evaluation of whole blood theophylline enzyme immunochromatography assay,

M. P. Habib, R. B. Schifman, B. Y . Shon, J. F. Fiastro, and S. C. Campbell, “Evaluation of whole blood theophylline enzyme immunochromatography assay,” Chest, vol. 92, no. 1, pp. 129–131, Jul. 1987. doi: 10.1378/chest.92.1.129

work page doi:10.1378/chest.92.1.129 1987

[4] [4]

Hansch, A

C. Hansch, A. Leo, and D. Hoekman, Exploring QSAR: Hydrophobic, Electronic, and Steric Constants, vol. 2. Washington, DC, USA: Amer. Chem. Soc., 1995

1995

[5] [5]

Caco-2 cell line standardization with pharmaceutical requirements and in vitro model suitability for permeability assays,

M. Kus, I. Ibragimow, and H. Piotrowska-Kempisty, “Caco-2 cell line standardization with pharmaceutical requirements and in vitro model suitability for permeability assays,” Pharmaceutics, vol. 15, no. 11, Art. no. 2523, Oct. 2023. doi: 10.3390/pharmaceutics15112523

work page doi:10.3390/pharmaceutics15112523 2023

[6] [6]

Optimization of physicochemical properties of theophylline by forming cocrystals with amino acids,

Q. Li, Y . Xie, Z. Wang et al. , “Optimization of physicochemical properties of theophylline by forming cocrystals with amino acids,” RSC Adv., vol. 14, no. 54, pp. 40006–40017, Nov. 2024. doi: 10.1039/d4ra06804a

work page doi:10.1039/d4ra06804a 2024

[7] [7]

Clinical pharmacodynamics of theophylline,

C. C. Peck, A. I. Nichols, J. Baker, L. L. Lenert, and D. Ezra, “Clinical pharmacodynamics of theophylline,” J. Allergy Clin. Immunol., vol. 76, no. 2, pp. 292– 297, Aug. 1985. doi: 10.1016/0091-6749(85)90644-x

work page doi:10.1016/0091-6749(85)90644-x 1985

[8] [8]

Metabolism of theophylline by cDNA-expressed human cytochromes P -450,

H. R. Ha, J. Chen, A. U. Freiburghaus, and F. Follath, “Metabolism of theophylline by cDNA-expressed human cytochromes P -450,” Br. J. Clin. Pharmacol., vol. 39, no. 3, pp. 321–326, Mar. 1995. doi: 10.1111/j.1365-2125.1995.tb04455.x

work page doi:10.1111/j.1365-2125.1995.tb04455.x 1995

[9] [9]

Abduljalil, I

K. Abduljalil, I. Gardner, and M. Jamei, “Application of a physiologically based pharmacokinetic approach to predict theophylline pharmacokinetics using virtual non- pregnant, pregnant, fetal, breast-feeding, and neonatal populations,” Front. Pediatr., vol. 10, Art. no. 840710, Mar. 2022. doi: 10.3389/fped.2022.840710

work page doi:10.3389/fped.2022.840710 2022