Post-Inference Model Chaining: A Problem Statement and a Model-based Bootstrap for Training End-to-End Digital Twins

Lukas Madenach, Timo Bailer, Tobias EilertarXiv 2026 · arXiv

TL;DR

Building "End-to-End Digital Twins" by chaining separate process models often leads to massive prediction biases and overly conservative, "inflated" uncertainty intervals. This paper identifies why the standard Monte Carlo approach fails in nonlinear systems and introduces a model-based bootstrap method that drastically improves accuracy and calibration for multi-stage industrial processes.

What Problem Does This Paper Address?

In industries like biopharmaceuticals, a "Digital Twin" is a virtual replica of a physical production line. Because these lines are complex and data is often scarce, engineers usually model each stage (or "unit operation") separately and then chain them together to predict the final product quality.

The industry standard for this is a Monte Carlo simulation (MCSim) approach: you take a random draw from the output distribution of Stage 1, plug it into Stage 2, and repeat until the end. However, the authors show that this "predictive-draw propagation" is statistically flawed. In nonlinear systems, it introduces a permanent bias that doesn't disappear even if you collect infinite data. Furthermore, it results in prediction intervals that are far too wide—sometimes 4 to 5 times larger than necessary—making the digital twin less useful for actual process control.

Methods

The authors propose a two-phase fixed-center studentized model-based bootstrap to replace the standard MCSim approach.

  1. Holistic Retraining: Instead of training Stage 2 on the noisy observed outputs of Stage 1, they train it on the fitted mean predictions of Stage 1. This helps mitigate the "error-in-variables" problem, where measurement noise in the predictors biases the regression coefficients.
  2. Sequential Bootstrap: To estimate uncertainty, they generate "surrogate" datasets based on the initial models. They refit the entire chain of models for each bootstrap iteration. Crucially, they forward the mean predictions through the bootstrap chain to capture the dependencies between model coefficients across different stages.
  3. Studentized Root: Finally, they use a studentized prediction root to calculate the final endpoint intervals, ensuring the uncertainty accounts for both parameter estimation error and future measurement noise.

Datasets

Because the goal was to compare model performance against an absolute "truth," the authors used simulated deterministic chains where the exact parameters were known:

  • Univariate Chains: 4-stage (LQLQ) and 8-stage (LQLQLQLQ) processes, where "L" represents a linear stage and "Q" represents a quadratic stage.
  • Bivariate Chain: A more complex 6-stage process included in the supplements to test the method's scalability to multiple variables.
  • Sample Sizes: They tested small-data scenarios typical of industrial R&D ($m = 7, 15, \text{ and } 30$ training runs).

Key Insights and Ideas

  • Nonlinearity is the "Silent Killer": In a simple linear chain, MCSim is unbiased but inefficient. However, as soon as a quadratic stage is introduced, MCSim creates a mean bias that persists asymptotically. The math simply doesn't "add up" when you propagate full distributions through nonlinear functions.
  • The Error-in-Variables Trap: When we chain models, the output of Stage 1 becomes the input for Stage 2. If Stage 1's output is noisy, standard regression at Stage 2 becomes biased. The authors' "fixed-center" approach effectively "denoises" the training data for subsequent stages.
  • Undefined Moments: A surprising finding was that in some MCSim configurations, the predicted mean or variance might not even mathematically exist (infinite moments), leading to highly unstable numerical results.

Key Results

The proposed bootstrap method outperformed the standard MCSim across every metric in the 8-stage (LQLQLQLQ) simulation at a sample size of $m=30$:

  • Bias Reduction: The relative mean bias for the bootstrap was between -3.16% and +2.65%, whereas the MCSim bias ranged from -45.0% to +34.1%.
  • Interval Precision: The bootstrap produced intervals that were only 1.16 to 1.42 times the width of the "ideal" oracle reference. In contrast, the MCSim intervals were 4.11 to 4.97 times wider than necessary.
  • Coverage Calibration: The bootstrap maintained coverage close to the nominal 90% target, while MCSim was often "too safe," providing 95% to 100% coverage at the cost of being wildly imprecise.

Limitations and Future Work

  • Error Distribution Assumptions: The current method assumes the model knows its own error distribution (e.g., Gaussian). While common in linear mixed-effects models, it might require adaptation for non-parametric models.
  • Process Variability: The study focused on measurement error. The authors note that incorporating "process variability" (actual physical fluctuations in the stage means) is a necessary next step for even more realistic digital twins.
  • Model Misspecification: The study assumes the functional form (linear or quadratic) is correctly chosen. In the real world, model-form error remains a challenge that the bootstrap might mitigate but not eliminate.

Summary updated

Abstract

End-to-end digital twins can be constructed by fitting separate stage models and connecting them after inference. We investigate the Frequentist prediction properties of a commonly used Monte Carlo simulation scheme that forwards complete predictive draws through such a chain when measured stage outputs contain independent normal errors, but the latent stage means are deterministic. Analytical two-stage examples show persistent asymptotic interval and coverage inflation and, with quadratic propagation, mean bias dependent on the parameters to be estimated. Simulations of four- and eight-stage linear-quadratic univariate chains with 7, 15 or 30 training runs demonstrate pronounced input-dependent distortions. A diagnostic using error-free latent training predictors shows that the effect is mainly not attributable to errors in variables, but to the MCSim approach even showing underestimation of coverage. We propose a model-based bootstrap: an initial chain forwards fitted means, bootstrap chains are refitted on fixed-center bootstrap data while forwarding their mean prediction, and a target-stage future error enters a studentized prediction root. Across the investigated scenarios, relative mean bias is substantially reduced, and coverage is closer to the nominal 90%. At sample size 30, the eight-stage bootstrap has mean width ratios of 1.16-1.42 relative to the known-parameter normal reference, compared with 4.11-4.97 for observed-to-observed MCSim. These results support improved endpoint prediction in the tested setting. Further investigations of the bootstrap scheme in the supplement for a bivariate 6-stage process show similar satisfactory results.
BibTeX
@misc{madenach2026post,
  author = {Lukas Madenach and Timo Bailer and Tobias Eilert},
  title = {{Post-Inference Model Chaining: A Problem Statement and a Model-based Bootstrap for Training End-to-End Digital Twins}},
  year = {2026},
  eprint = {2611.05225},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.05225},
}
Figure