Meta-Inverse PINN for High-Dim ODEs

• MI-PINN is a meta-learning framework that first learns a reusable, physics-aware representation across related ODE tasks before performing task-specific inverse inference via a closed-form pseudo-inverse.
• It improves parameter recovery and stability by decoupling shared representation learning from the optimization of task-specific variables, enhancing conditioning and sample efficiency.
• Adaptive clustering with multi-branch neural networks addresses multi-scale dynamics, as demonstrated in accurate PBPK model parameter inference and missing mechanism reconstruction.

Searching arXiv for the primary paper and a small set of closely related PINN/meta-learning references to ground the article. Meta-Inverse Physics-Informed Neural Network (MI-PINN) is a two-stage meta-learning formulation for inverse problems in high-dimensional, tightly coupled ordinary differential equation systems under partial observability and sparse data. It reformulates inverse modeling by first learning a reusable, physics-aware representation across a family of related tasks and then performing task-specific inverse inference while keeping that representation fixed and recomputing the final layer in closed form through a physics-informed pseudo-inverse. In "Meta-Inverse Physics-Informed Neural Networks for High-Dimensional Ordinary Differential Equations" (Wei et al., 5 May 2026), the method is demonstrated on whole-body physiologically based pharmacokinetic (PBPK) models with up to 33 coupled ODEs for paracetamol and theophylline under intravenous and oral dosing, where it is used for masked kinetic-parameter recovery and reconstruction of missing mechanistic terms.

1. Problem formulation and inverse setting

MI-PINN is designed for high-dimensional ODE systems of the form

$$\dot{\mathbf{x}}(t)=f\big(\mathbf{x}(t),\mathbf{u}(t),t;\boldsymbol{\theta}\big), \qquad \mathbf{x}(0)=\mathbf{x}_0,$$

where $\mathbf{x}(t)\in\mathbb{R}^S$ denotes the state vector, $\mathbf{u}(t)$ denotes external inputs such as dosing, and $\boldsymbol{\theta}$ collects physical parameters such as kinetic rates and partition coefficients. In the reported PBPK applications, $S=22$ for intravenous paracetamol and $S=33$ for oral paracetamol.

The inverse problem is posed under partial observability. Observations are available only on a subset of states, typically venous blood concentration, according to

$$\mathbf{y}(t_j)=h\big(\mathbf{x}(t_j)\big)+\boldsymbol{\epsilon}_j, \qquad \boldsymbol{\epsilon}_j\sim\mathcal{N}(0,\Sigma),$$

with $h:\mathbb{R}^S\to\mathbb{R}^{S_{\text{obs}}}$ selecting the observed channels. In the paracetamol intravenous setting, the observation set consists of 10 temporal measurements on a single observable channel. The task family $\mathcal{T}_i$ comprises PBPK systems with shared mechanistic structure but different parameter configurations or different missing-term formulations. Examples of unknown parameter vectors are $\boldsymbol{\theta}_i=[K_{m,P},V_{\max,P}]$ for paracetamol UGT2B15 and $\mathbf{x}(t)\in\mathbb{R}^S$0 for theophylline CYP1A2-mediated DMU formation (Wei et al., 5 May 2026).

Initial conditions and physical constraints are integral to the formulation. Intravenous dosing introduces infusion into blood over time, oral dosing initializes drug in stomach fluid or lumen, and other compartments start at zero. Concentrations are constrained by non-negativity and unit consistency. A central identifiability issue is that, under sparse observations, recovery may depend more strongly on the catalytic-efficiency ratio $\mathbf{x}(t)\in\mathbb{R}^S$1 than on the individual parameters. This is important for interpreting inverse results: accurate trajectory recovery does not by itself imply full structural identifiability of every kinetic constant.

2. Two-stage meta-learning formulation

The MI-PINN architecture separates shared representation learning from task-specific inverse inference. The shared, physics-aware representation parameters are denoted $\mathbf{x}(t)\in\mathbb{R}^S$2, one parameter set per cluster-specific branch. Each branch produces features $\mathbf{x}(t)\in\mathbb{R}^S$3 for task $\mathbf{x}(t)\in\mathbb{R}^S$4, where $\mathbf{x}(t)\in\mathbb{R}^S$5 indexes the task configuration. The final layer parameters are $\mathbf{x}(t)\in\mathbb{R}^S$6, one vector per state. Task-specific inverse variables are $\mathbf{x}(t)\in\mathbb{R}^S$7, which may include masked kinetic parameters or parameters of auxiliary neural approximators for missing mechanisms.

During Stage 1, MI-PINN meta-trains the shared representation $\mathbf{x}(t)\in\mathbb{R}^S$8 by minimizing a per-task PINN objective combining initial-condition, ODE-residual, and data terms: $\mathbf{x}(t)\in\mathbb{R}^S$9 With collocation sets $\mathbf{u}(t)$0, these terms are

$\mathbf{u}(t)$1

$\mathbf{u}(t)$2

$\mathbf{u}(t)$3

The predicted state $\mathbf{u}(t)$4 assigned to cluster $\mathbf{u}(t)$5 is represented as

$\mathbf{u}(t)$6

Given $\mathbf{u}(t)$7, the final layer is not updated by gradient descent. Instead, MI-PINN assembles a physics-informed feature matrix $\mathbf{u}(t)$8 and target vector $\mathbf{u}(t)$9 encoding linearized initial-condition, ODE, and data constraints, and solves

$\boldsymbol{\theta}$0

with Tikhonov regularization $\boldsymbol{\theta}$1. For nonlinearities, a lagged-coefficient freezing method linearizes the constraints at each iteration. The meta-objective is

$\boldsymbol{\theta}$2

Stage 2 performs decoupled inverse inference. The learned representation $\boldsymbol{\theta}$3 is fixed, and only the task-specific unknowns $\boldsymbol{\theta}$4 are optimized: $\boldsymbol{\theta}$5 where

$\boldsymbol{\theta}$6

The central dimensionality argument is explicit: MI-PINN transforms a joint search over network weights and task unknowns of dimension $\boldsymbol{\theta}$7 into a search only over task unknowns of dimension $\boldsymbol{\theta}$8, while $\boldsymbol{\theta}$9 is solved optimally in closed form under a fixed shared basis. The reported interpretation is that this improves conditioning, identifiability, and sample efficiency (Wei et al., 5 May 2026).

3. Adaptive clustering-based multi-branch representation

To handle multi-scale dynamics in high-dimensional ODE systems, MI-PINN introduces an adaptive clustering-based multi-branch learning scheme. The clustering procedure begins by collecting a trajectory vector $S=22$0 for each state variable $S=22$1 over a preliminary simulation or representative task family. Pairwise correlation and distance are then defined as

$S=22$2

Agglomerative hierarchical clustering on the distance matrix $S=22$3 yields $S=22$4 clusters, and each state $S=22$5 is assigned to a cluster $S=22$6.

Each cluster $S=22$7 has a branch neural network $S=22$8 producing features $S=22$9. Under hard assignment,

$S=33$0

The final layer remains physics-integrated: the matrix $S=33$1 is assembled from linearized initial-condition, ODE, and data constraints, and $S=33$2 is again obtained from

$S=33$3

The stated rationale for the multi-branch construction is that branches specialize in the time scales and dynamic patterns of their assigned states, thereby countering spectral bias and gradient pathologies in multi-scale ODEs. In the PBPK experiments, $S=33$4 worked well for intravenous settings and $S=33$5 for oral settings. The reported branch architecture is $S=33$6, with sine activations used to represent multi-scale dynamics. A plausible implication is that the multi-branch decomposition is not merely a capacity increase; it imposes a state-grouped inductive bias aligned with cross-compartment temporal heterogeneity (Wei et al., 5 May 2026).

4. PBPK systems, tasks, and training protocol

The principal applications are whole-body PBPK models for paracetamol and theophylline. The compartment structure includes arterial and venous blood and the organs brain, lung, heart, stomach, spleen, pancreas, adipose, skin, muscle, bone, and rest-of-body, with liver and kidney as primary elimination sites. Oral dosing extends the model with gut luminal and wall compartments for stomach fluid, duodenum, jejunum, ileum, and large intestine, together with dissolution and transit.

The governing PBPK equations include perfusion-limited organ mass balance,

$S=33$7

with the lungs swapping arterial and venous concentrations accordingly; the intestinal absorption rate constant,

$S=33$8

the liver metabolism equation with summed enzyme-mediated Michaelis–Menten terms scaled by IVIVE; and the renal clearance equation

$S=33$9

The inverse tasks include parameter inference and missing-mechanism reconstruction. For paracetamol, the masked unknowns are UGT2B15 $$\mathbf{y}(t_j)=h\big(\mathbf{x}(t_j)\big)+\boldsymbol{\epsilon}_j, \qquad \boldsymbol{\epsilon}_j\sim\mathcal{N}(0,\Sigma),$$0; for theophylline, they are CYP1A2 $$\mathbf{y}(t_j)=h\big(\mathbf{x}(t_j)\big)+\boldsymbol{\epsilon}_j, \qquad \boldsymbol{\epsilon}_j\sim\mathcal{N}(0,\Sigma),$$1 for DMU formation. In the missing-mechanism setting, the entire UGT2B15 Michaelis–Menten term in the paracetamol liver ODE, including IVIVE scaling, is treated as unknown and learned by an auxiliary neural network. The reported system dimensions are 22 coupled ODEs for intravenous scenarios, of which 20 are linear and 2 nonlinear, and 33 coupled ODEs for oral scenarios, of which 21 are linear and 12 nonlinear (Wei et al., 5 May 2026).

Training follows the two-stage scheme. Stage 1 uses 20 tasks with differing $$\mathbf{y}(t_j)=h\big(\mathbf{x}(t_j)\big)+\boldsymbol{\epsilon}_j, \qquad \boldsymbol{\epsilon}_j\sim\mathcal{N}(0,\Sigma),$$2 configurations or missing-term formulations, 200 temporal collocation points per task, Adam with learning rate $$\mathbf{y}(t_j)=h\big(\mathbf{x}(t_j)\big)+\boldsymbol{\epsilon}_j, \qquad \boldsymbol{\epsilon}_j\sim\mathcal{N}(0,\Sigma),$$3, and a physics-informed pseudo-inverse for the final layer at every iteration. Stage 2 fixes $$\mathbf{y}(t_j)=h\big(\mathbf{x}(t_j)\big)+\boldsymbol{\epsilon}_j, \qquad \boldsymbol{\epsilon}_j\sim\mathcal{N}(0,\Sigma),$$4 and optimizes $$\mathbf{y}(t_j)=h\big(\mathbf{x}(t_j)\big)+\boldsymbol{\epsilon}_j, \qquad \boldsymbol{\epsilon}_j\sim\mathcal{N}(0,\Sigma),$$5 using Adam; for missing terms, the auxiliary network is $$\mathbf{y}(t_j)=h\big(\mathbf{x}(t_j)\big)+\boldsymbol{\epsilon}_j, \qquad \boldsymbol{\epsilon}_j\sim\mathcal{N}(0,\Sigma),$$6. For paracetamol intravenous experiments, the observed data are 10 aggregated venous blood observations from a clinical study. For theophylline intravenous experiments, the observed data are 8 individual venous blood profiles with doses ranging from 194 mg to 383 mg over 40 minutes. The stopping criterion is convergence of validation loss, and no explicit curriculum is required.

5. Empirical behavior and comparative results

The reported empirical results span parameter inference, missing-mechanism reconstruction, comparison against solver-based optimization, comparison against conventional joint-optimization PINNs, and a multi-branch ablation. The numerical outcomes are summarized below.

Scenario	Observations / system	Reported outcome
IV paracetamol parameter inference	10 venous blood points; 22 ODEs	Trajectory MSE across 22 states: 0.96
Oral paracetamol parameter inference	1000 mg tablet; 33 ODEs	Accurate predictions across all 33 states
IV paracetamol missing mechanism	Entire UGT2B15 term masked; 22 ODEs	Trajectory MSE across 22 states: $$\mathbf{y}(t_j)=h\big(\mathbf{x}(t_j)\big)+\boldsymbol{\epsilon}_j, \qquad \boldsymbol{\epsilon}_j\sim\mathcal{N}(0,\Sigma),$$7
IV theophylline parameter inference	8 IV profiles; 22 ODEs	MSE across 22 states: $$\mathbf{y}(t_j)=h\big(\mathbf{x}(t_j)\big)+\boldsymbol{\epsilon}_j, \qquad \boldsymbol{\epsilon}_j\sim\mathcal{N}(0,\Sigma),$$8
Single- vs multi-branch ablation	IV theophylline forward modeling	$$\mathbf{y}(t_j)=h\big(\mathbf{x}(t_j)\big)+\boldsymbol{\epsilon}_j, \qquad \boldsymbol{\epsilon}_j\sim\mathcal{N}(0,\Sigma),$$9 vs $h:\mathbb{R}^S\to\mathbb{R}^{S_{\text{obs}}}$0

For intravenous paracetamol parameter inference after a 2 h infusion at 20 mg/kg, the ground truth is $h:\mathbb{R}^S\to\mathbb{R}^{S_{\text{obs}}}$1 $h:\mathbb{R}^S\to\mathbb{R}^{S_{\text{obs}}}$2, $h:\mathbb{R}^S\to\mathbb{R}^{S_{\text{obs}}}$3 $h:\mathbb{R}^S\to\mathbb{R}^{S_{\text{obs}}}$4, and ratio $h:\mathbb{R}^S\to\mathbb{R}^{S_{\text{obs}}}$5. Trust-region reflective optimization in MATLAB yields $h:\mathbb{R}^S\to\mathbb{R}^{S_{\text{obs}}}$6, $h:\mathbb{R}^S\to\mathbb{R}^{S_{\text{obs}}}$7, and ratio $h:\mathbb{R}^S\to\mathbb{R}^{S_{\text{obs}}}$8. MI-PINN yields $h:\mathbb{R}^S\to\mathbb{R}^{S_{\text{obs}}}$9, $\mathcal{T}_i$0, and ratio $\mathcal{T}_i$1. The reported interpretation is that MI-PINN recovers both trajectory and parameters more accurately than solver-based optimization despite sparse single-channel data.

For oral paracetamol parameter inference with a 1000 mg tablet, the ground truth is $\mathcal{T}_i$2, $\mathcal{T}_i$3, ratio $\mathcal{T}_i$4. Trust-region reflective optimization yields $\mathcal{T}_i$5, $\mathcal{T}_i$6, ratio $\mathcal{T}_i$7. MI-PINN yields $\mathcal{T}_i$8, $\mathcal{T}_i$9, ratio $\boldsymbol{\theta}_i=[K_{m,P},V_{\max,P}]$0. The result is presented as accurate prediction across all 33 states with improved consistency of the catalytic-efficiency ratio.

For intravenous paracetamol missing-mechanism discovery, MI-PINN learns the time course of the fully masked UGT2B15 term with an auxiliary neural network. Dimensional analysis and symbolic regression on the learned outputs recover the near-correct functional form

$\boldsymbol{\theta}_i=[K_{m,P},V_{\max,P}]$1

The omitted small-denominator terms are reported as negligible under the linear regime $\boldsymbol{\theta}_i=[K_{m,P},V_{\max,P}]$2, which explains why trajectories remain accurate even though the reconstructed expression is not fully identical to the original mechanistic term.

For intravenous theophylline parameter inference, based on 8 individual IV profiles with doses from 194 mg to 383 mg over 40 min, the ground truth is $\boldsymbol{\theta}_i=[K_{m,P},V_{\max,P}]$3 $\boldsymbol{\theta}_i=[K_{m,P},V_{\max,P}]$4 and $\boldsymbol{\theta}_i=[K_{m,P},V_{\max,P}]$5 $\boldsymbol{\theta}_i=[K_{m,P},V_{\max,P}]$6. Trust-region reflective optimization yields $\boldsymbol{\theta}_i=[K_{m,P},V_{\max,P}]$7, $\boldsymbol{\theta}_i=[K_{m,P},V_{\max,P}]$8, whereas MI-PINN yields $\boldsymbol{\theta}_i=[K_{m,P},V_{\max,P}]$9, $\mathbf{x}(t)\in\mathbb{R}^S$00. Against a conventional jointly optimized PINN on the same task, the reported baseline estimates are $\mathbf{x}(t)\in\mathbb{R}^S$01 and $\mathbf{x}(t)\in\mathbb{R}^S$02 with poor trajectories, while MI-PINN attains near-ground-truth parameters and MSE $\mathbf{x}(t)\in\mathbb{R}^S$03. The reported error reduction reaches up to two orders of magnitude. A single forward run for the 22-ODE theophylline system requires 14.66 s with SciPy odeint and 4.17 s with MI-PINN, corresponding to an approximately $\mathbf{x}(t)\in\mathbb{R}^S$04 speedup (Wei et al., 5 May 2026).

6. Scalability, limitations, and interpretive issues

The computational bottleneck in the final-layer solve is a regularized least-squares problem with feature dimension $\mathbf{x}(t)\in\mathbb{R}^S$05 and constraint count $\mathbf{x}(t)\in\mathbb{R}^S$06, with cost roughly $\mathbf{x}(t)\in\mathbb{R}^S$07 for forming $\mathbf{x}(t)\in\mathbb{R}^S$08 and inverting it; Cholesky or QR can be used. Multi-branch modeling increases parameter count linearly with the number of clusters $\mathbf{x}(t)\in\mathbb{R}^S$09, but the reported trade-off is reduced interference across time scales and improved runtime because $\mathbf{x}(t)\in\mathbb{R}^S$10 updates are gradient-free and conditioning is better. The method is shown to scale to $\mathbf{x}(t)\in\mathbb{R}^S$11 ODEs with improved stability.

Several limitations are explicit. MI-PINN requires a family of related tasks for meta-pretraining, and large domain shift may degrade transfer. Partial observability can cause identifiability problems, particularly for separating individual $\mathbf{x}(t)\in\mathbb{R}^S$12 and $\mathbf{x}(t)\in\mathbb{R}^S$13 values from their ratio; MI-PINN mitigates but does not eliminate this non-identifiability. Missing-mechanism discovery depends on auxiliary-network capacity and on training data spanning the relevant dynamical regimes; saturated or nonlinear regimes are needed to recover denominators faithfully. These caveats qualify a common misconception that successful trajectory fitting necessarily implies exact mechanistic recovery.

The practical guidance in the reported study reflects these constraints. Correlation-based clustering on preliminary trajectories is used to group states, with $\mathbf{x}(t)\in\mathbb{R}^S$14 for intravenous and $\mathbf{x}(t)\in\mathbb{R}^S$15 for oral PBPK systems. Loss weighting should begin with $\mathbf{x}(t)\in\mathbb{R}^S$16 under sparse data, and the Tikhonov weight $\mathbf{x}(t)\in\mathbb{R}^S$17 is tuned to stabilize the pseudo-inverse. Two hundred temporal collocation points per task are reported as sufficient, with higher density suggested when trajectories exhibit sharp transients or stiff behavior. The representation task set should span plausible parameter ranges; in the reported implementation, 20 task configurations are used.

Theoretical remarks in the study are stated informally. Fixing the learned representation $\mathbf{x}(t)\in\mathbb{R}^S$18 constrains the inverse search to task-specific variables $\mathbf{x}(t)\in\mathbb{R}^S$19, thereby restricting the hypothesis space to a stable, physics-aware basis. Likewise, solving the final layer by physics-informed pseudo-inverse eliminates gradient updates for $\mathbf{x}(t)\in\mathbb{R}^S$20, which is argued to improve stability and conditioning under multi-scale dynamics and non-convex optimization landscapes. This suggests that MI-PINN should be understood not as a generic replacement for PINNs, but as a specific reorganization of inverse modeling that is most effective when a reusable cross-task basis can be learned and when physics constraints can be embedded into the closed-form final-layer solve (Wei et al., 5 May 2026).