📊 statistics

Non-Parametric Model Calibration with Stochastic Control Parameters

This paper presents a non-parametric method for calibrating computer models with stochastic inputs that generates a distributional estimate consistent with observed field data while preserving the known distributions of specified control parameters.

Original authors: Akshay Prasadan, Samopriya Basu, Faezeh Yazdi, Derek Bingham, Donald Estep

Published 2026-07-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Akshay Prasadan, Samopriya Basu, Faezeh Yazdi, Derek Bingham, Donald Estep

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery, but the clues you have are a bit messy. You have a computer model—a digital twin of the real world—that acts like a magic machine. You feed it some ingredients (inputs), and it spits out a result (output). Usually, scientists want to work backward: they see the result in the real world and try to figure out exactly what ingredients were used to create it. This is called "calibration."

However, life is rarely that simple. Sometimes, the ingredients you put into your machine aren't just one thing; they are a mix of two types. First, there are the "secret spices" (calibration parameters) that you don't know anything about. You have no idea how much of them were used, or how they are distributed. Second, there are the "known variables" (control parameters). Maybe you know the oven was set to a specific temperature range, or you know the wind was blowing from the north with a predictable pattern. These are the control parameters. The problem is, most detective work assumes you know nothing about the ingredients, or that you know everything. But what if you know the recipe for the wind, but not the secret spices? That is the tricky corner of science this paper explores. It asks: How do we figure out the secret spices without messing up the known recipe for the wind?

The authors of this paper, Akshay Prasadan and his team, have built a new method to solve this specific puzzle. They are working in the field of statistics and computer modeling, specifically dealing with "inverse problems" (working backward from results to causes). Their big idea is to create a mathematical "safety net" that ensures their solution for the secret spices doesn't accidentally change the known distribution of the control parameters.

Here is how their method works, using a playful analogy. Imagine you are trying to guess the flavor of a mystery soup (the secret spices) by tasting bowls of soup that were made in different kitchens (the control parameters). You know that in every kitchen, the amount of salt (a control parameter) follows a very specific, well-known pattern—maybe half the kitchens use a pinch, and half use a teaspoon. But you don't know how much pepper (the secret calibration parameter) was added.

If you just use a standard detective method, you might guess a pepper distribution that fits the soup taste, but in doing so, you might accidentally guess that the salt distribution changed too. Maybe your guess suggests that all the kitchens suddenly started using a cup of salt! That would be scientifically wrong because you know the salt pattern didn't change. The authors' new method prevents this. They use a technique called "optimal transport," which is like a super-smart logistics planner. Imagine you have a pile of salt bags (the known distribution) and a pile of mystery soup bowls (the observed data). The planner figures out the most efficient way to match them up so that the salt bags end up exactly where they are supposed to be, while simultaneously figuring out the pepper distribution that explains the soup's taste.

The paper demonstrates this with two main experiments. First, they used a simple math model (a quadratic equation) where they could test their method perfectly. They found that when they had data where the salt and soup were recorded together (paired data), their method worked great. But the real magic happened when they only had the soup taste and a separate list of salt amounts (unpaired data). In this harder scenario, they used their "logistics planner" (optimal transport) to create a fake, but mathematically perfect, list of paired data. They then ran their calibration on this fake list. The result? Their method successfully recovered the correct pepper distribution while strictly keeping the salt distribution exactly as it was supposed to be.

They also tested this on a more complex, real-world scenario: simulating heat moving through a thin metal plate. Here, the "secret spices" were the metal's ability to conduct heat, and the "known variable" was the position of the heat source, which wiggled around in a known pattern. Because solving the heat equations on a computer takes a long time, they trained a fast "surrogate" model (a machine learning shortcut) to act as the computer model. Even with this shortcut, their method held up. They showed that they could estimate the heat-conducting properties of the metal accurately without distorting the known pattern of the heat source's movement.

The authors are careful to note that while their method works beautifully in these simulations, it relies on the computer model behaving nicely (mathematically speaking). They also point out that using a fast surrogate model (the machine learning shortcut) introduces a tiny bit of extra error, which is a trade-off for speed. However, their core finding is robust: by augmenting the model and using optimal transport, you can calibrate the unknown parts of a system without breaking the rules of the known parts. This is a significant step forward for scientists who need to trust their models while dealing with messy, real-world data where some things are known and others are a complete mystery.

Technical Summary: Non-Parametric Model Calibration with Stochastic Control Parameters

Problem Statement
The paper addresses the Statistical Calibration Problem (SCP) for computer models where inputs are stochastic. In standard SCP settings, all stochastic inputs are treated as calibration parameters with unknown distributions, and the goal is to find an input distribution whose pushforward through the model matches observed field data. However, many physical systems involve two distinct types of stochastic inputs:

Calibration Parameters ( $\Lambda$ ): Unknown parameters (e.g., material properties) with unknown distributions that must be inferred.
Control Parameters ( $U$ ): Parameters representing experimental conditions (e.g., heat source position, seasonal temperature) that vary stochastically but have a known distribution.

The core challenge is to estimate the distribution of the calibration parameters such that the resulting model output is consistent with field data, while simultaneously ensuring that the marginal distribution of the control parameters in the solution matches the known, specified distribution. A naive application of existing non-parametric calibration techniques (which treat all inputs as unknown) fails to preserve the known marginal of the control parameters, leading to scientifically implausible estimates.

Methodology
The authors propose a constrained calibration framework that extends the non-parametric approach based on the disintegration of measures and Bayesian inference (specifically referencing the work of [9]). The methodology proceeds in three main stages:

Augmentation of the Output Space:
To enforce the constraint on the control parameters, the authors define an augmented map $\tilde{Q}: \Lambda \times U \to D \times U$ where $\tilde{Q}(\lambda, u) = (Q(\lambda, u), u)$ . Theorem 1 establishes that finding a probability measure on the input space whose pushforward under $\tilde{Q}$ matches the target joint distribution is equivalent to solving the constrained SCP (matching both the output distribution $PD$ and the control marginal $P^t_U$ ).
Handling Paired Data:
In the scenario where field observations are recorded alongside the specific control parameter values for each trial (paired data), the joint distribution of the augmented output is directly observable. In this case, the problem reduces to an unconstrained SCP on the augmented space, solvable directly using the existing Algorithm 1 from [9] with the augmented map and a prior that respects the known marginal of $U$ .
Handling Unpaired Data via Optimal Transport:
In the more challenging scenario where control parameter values are not recorded with specific field observations (unpaired data), the joint distribution of the augmented output is unknown, though its marginals ($PD$ and $P^t_U$ ) are known. The authors employ Optimal Transport (OT) to construct a surrogate joint distribution.
- They formulate an entropic regularized Kantorovich problem to find a joint measure $\hat{\Pi}$ on $D \times U$ with the correct marginals that minimizes a cost function.
- Crucially, the cost function $c(q, u)$ is defined as the expected squared error between the observed output $q$ and the model output $Q(\lambda, u)$ , integrated over the prior of $\Lambda$ . This incorporates the geometry of the computer model into the transport plan.
- The Sinkhorn-Knopp (SK) algorithm is used to solve for $\hat{\Pi}$ .
- Synthetic paired data is generated from $\hat{\Pi}$ , which is then fed into the augmented SCP algorithm (Algorithm 2) to estimate the final input distribution.

Key Contributions

Constrained Calibration Framework: The paper introduces a rigorous method for calibrating computer models where some stochastic inputs have known distributions, ensuring the solution respects these physical constraints.
Augmentation Strategy: It demonstrates that augmenting the output space with the control parameters transforms the constrained problem into a standard unconstrained problem on a higher-dimensional space.
Optimal Transport Integration: For unpaired data, the paper proposes a novel use of optimal transport with a model-dependent cost function to reconstruct the necessary joint distribution of outputs and control parameters, enabling calibration without paired observations.
Algorithmic Implementation: The authors provide specific algorithms (Algorithms 2 and 3) and demonstrate their implementation using histogram estimators and the Sinkhorn-Knopp algorithm.

Results
The authors validate the method through two numerical experiments:

Quadratic Model: A simple quadratic model with additive noise serves as a testbed. The results show that the unconstrained SCP solution fails to recover the known bimodal distribution of the control parameter, whereas the proposed constrained method successfully recovers the correct marginal. The unpaired data approach yields results nearly identical to the paired data benchmark.
Heat Transfer in Metal Plates: A more complex application involving the steady-state solution of a 2D heat equation. The calibration parameters are convection and diffusion coefficients, while the control parameter is the vertical position of the heat source. The authors utilize a surrogate model (XGBoost) trained on PDE solutions to accelerate the computationally intensive SK and importance sampling steps.
- The results demonstrate that the constrained method produces input distributions consistent with the observed temperature data and the known Beta distribution of the heat source position.
- Diagnostic checks confirm that the pushforward of the estimated solution matches the observed data distribution and that the marginal of the control parameter matches the specified distribution.
- The study notes that while the unpaired method is robust, the contours in the marginal solutions can be less distinct than in the paired case, particularly for parameters with minimal impact on the output (e.g., convection coefficients).

Significance and Claims
The paper claims to extend the range of solvable calibration problems by addressing the "constrained" case where control parameters have known distributions. The authors emphasize that a scientifically plausible calibration must preserve well-established distributional properties of experimental conditions.

The significance of the work lies in:

Scientific Plausibility: Preventing calibration estimates from distorting known physical processes (e.g., ensuring a heat source's position distribution remains physically accurate).
Robustness to Data Limitations: The ability to handle unpaired data via optimal transport makes the method applicable to real-world scenarios where experimental metadata (control parameters) may not be perfectly synchronized with field measurements.
Theoretical Foundation: The method builds on the disintegration of measures, linking non-parametric calibration to Bayesian inference and providing a theoretical guarantee that the solution satisfies both the data consistency and marginal constraints.

The authors remain modest regarding future implications, noting that while the stochastic results may offer insights into deterministic calibration (via limits to Dirac delta measures), further investigation is required. They also acknowledge that the use of surrogate models introduces statistical error and potential violations of regularity conditions, warranting future theoretical work.

More like this