⚡ electrical engineering

VIBES -- A Two-Stage Scalable Bayesian Uncertainty Quantification Framework: Application to a Biomass Valorization Process

This paper introduces VIBES, a scalable two-stage Bayesian uncertainty quantification framework that combines Sobol global sensitivity analysis for dimensionality reduction with variational inference to significantly reduce predictive uncertainty in complex biomass valorization processes.

Original authors: Poulomi Das, Angan Mukherjee, Debangsu Bhattacharyya

Published 2026-07-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Poulomi Das, Angan Mukherjee, Debangsu Bhattacharyya

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 trying to bake the perfect loaf of bread, but you have a recipe with 20 different ingredients and settings (like oven temperature, flour type, yeast amount, humidity, etc.). You want to know exactly how much each of these factors affects the final taste and texture. However, testing every single combination of these 20 factors would take a lifetime and cost a fortune. This is the problem scientists face when trying to predict how complex chemical factories will behave.

This paper introduces a smart, two-step strategy called VIBES (Variational Inference-based Bayesian Estimation with Sobol screening) to solve this problem. Think of VIBES as a "smart filter" and a "fast-forward button" for scientific prediction.

Here is how it works, broken down into simple steps:

The Problem: Too Many Guesses

In chemical engineering, models are like giant, complicated recipes. They have many "knobs" (variables) you can turn, such as how fast you pump liquid, how hot you heat it, or how much raw material costs. Because we don't know the exact settings for every knob, there is "uncertainty." If you try to figure out the uncertainty for all 20 knobs at once, the computer gets overwhelmed, like a chef trying to taste-test 10,000 different bread recipes simultaneously.

Step 1: The "Sobol" Filter (Finding the Important Knobs)

The first stage of VIBES is like a detective looking for the most important clues. It uses a method called Sobol Global Sensitivity Analysis.

The Analogy: Imagine you have a car with 20 different parts (engine, tires, radio, seats, paint, etc.). You want to know which parts matter most for how fast the car goes. Instead of testing every single part, you run a simulation that tells you: "Actually, the engine and the tires matter 99% of the time. The radio and the paint color barely matter at all."
What VIBES did: It ran thousands of simulations on a bio-adhesive factory (a process that turns wood waste and soy protein into glue). It found that out of 20 uncertain factors, only 5 were actually driving the results. These were:
1. How much wood pulp (kraft lignin) flows in.
2. The price of soy protein.
3. The price of wood pulp.
4. Two specific chemical reaction speeds (activation energies).

By ignoring the other 15 "noise" factors, the problem became much smaller and easier to solve.

Step 2: The "Variational Inference" Fast-Forward (Learning from Data)

Once the team knew which 5 knobs mattered, they used the second stage: Variational Inference (VI).

The Analogy: Traditional methods of guessing the right settings are like trying to find a needle in a haystack by checking every single piece of hay one by one (slow and expensive). Variational Inference is like using a metal detector that scans the whole haystack at once and points you directly to the needle. It turns the guessing game into a math optimization problem, finding the "best fit" settings much faster.
The Twist: Since the factory model is still too complex for the computer to "read" directly, the team built a surrogate model (a simple, fast AI approximation) that mimics the factory. They used this AI to quickly learn the best settings for those 5 important knobs.

The Result: Sharper Predictions

After using VIBES, the team looked at their predictions for two things:

Process Performance: How much "D-lignin" (a key ingredient) they could produce.
Economic Performance: The "Levelized Cost of Adhesive" (how much it costs to make the glue).

The Magic Number: By using this two-stage method, they reduced the uncertainty in their predictions by more than 80%.

Before VIBES, their cost predictions were like a wide, fuzzy cloud of possibilities.
After VIBES, the cloud shrank into a tight, clear line.

They tested this in two ways:

Single Target: Giving the computer one specific goal (e.g., "The cost must be $0.785").
Multiple Targets: Giving the computer a range of data points (e.g., "The cost has been observed to be between $0.78 and $0.79").

Both methods worked, but using multiple data points (Approach 2) gave even tighter, more reliable results.

The "Python-Aspen" Bridge

To make this work, the researchers built a custom bridge between two software worlds:

Aspen Plus: The heavy-duty industrial software used to simulate the chemical factory (the "high-fidelity" model).
Python: The flexible coding language used to do the smart filtering and fast learning.

This bridge allowed them to automatically run the factory simulations, feed the results into the smart filter, and get the answers back without a human having to click buttons thousands of times.

Summary

The paper claims that VIBES is a powerful tool that:

Filters out the noise: It identifies the few variables that actually matter (reducing 20 variables down to 5).
Speeds up the math: It uses fast optimization techniques instead of slow, brute-force guessing.
Sharpens the picture: It reduces the uncertainty in predicting costs and production rates by over 80%, even when working with complex, high-dimensional chemical processes.

The authors demonstrated this specifically on a process to make bio-adhesives from wood waste and soy, proving that this method can help engineers design better, more reliable, and more cost-effective green technologies.

Technical Summary: VIBES – A Two-Stage Scalable Bayesian Uncertainty Quantification Framework

Problem Statement
High-dimensional process models used for chemical process design, techno-economic analysis (TEA), and sustainability assessment often suffer from significant uncertainties in kinetic, design/operational, and economic parameters. While uncertainty quantification (UQ) is critical for reliable decision-making, traditional sampling-based Bayesian methods (e.g., Markov chain Monte Carlo) become computationally intractable when applied to nonlinear, high-dimensional systems requiring repeated evaluations of rigorous process simulators. Furthermore, existing approaches often rely on local sensitivity analysis for dimensionality reduction, which may fail to capture global nonlinearities and interactions, or they utilize simplified surrogate models that lack the fidelity of first-principles simulations. There is a specific need for a scalable framework that can simultaneously identify dominant uncertainty sources and perform posterior estimation without sacrificing computational efficiency or model fidelity.

Methodology: The VIBES Framework
The paper proposes VIBES (Variational Inference-based Bayesian Estimation with Sobol screening), a two-stage scalable framework designed to address these challenges:

Stage 1: Sobol Global Sensitivity Analysis (GSA) for Dimensionality Reduction:
- The framework first performs a variance-based Sobol GSA on the full uncertain input space (including kinetic, design, and economic variables).
- Using total-order Sobol indices, the method identifies the subset of variables and parameters that govern the majority of the variance in process and economic outputs.
- This step reduces the dimensionality of the subsequent Bayesian inference problem, focusing only on the most influential inputs while capturing nonlinear effects and interactions.
Stage 2: Variational Inference (VI) for Bayesian Estimation:
- Bayesian inference is performed exclusively on the reduced-dimensional space identified in Stage 1.
- Instead of computationally expensive sampling, VI reformulates posterior estimation as an optimization problem. It approximates the true posterior distribution using a parameterized variational distribution (specifically, a diagonal Gaussian) by minimizing the Kullback-Leibler (KL) divergence.
- To ensure scalability, the VI optimization utilizes a differentiable surrogate model (a feedforward neural network) trained on the Sobol-sampled data. This allows for the use of stochastic gradient descent (SGD) and automatic differentiation (via JAX) to optimize variational parameters efficiently.
- A "warm-start" strategy is employed, initializing the variational parameters using the best-performing samples from the initial Sobol screening stage to accelerate convergence.
Posterior Predictive Uncertainty Propagation:
- Once the posterior distributions of the reduced-space variables are estimated, they are propagated through the high-fidelity process model (not the surrogate) to quantify the final predictive uncertainty.
- Non-selected variables are fixed at their base-case values during this final propagation step.

Implementation and Case Study
The framework was implemented using a multi-software integration platform connecting Python (for GSA, VI, and optimization) with Aspen Plus (for rigorous process simulation and economic evaluation). A Windows Component Object Model (COM) interface facilitated automated sampling, simulation, and data extraction.

The methodology was demonstrated on a novel bioadhesive production process involving the base-catalyzed depolymerization (BCD) of kraft lignin followed by crosslinking with soy protein isolate (SPI). The study focused on two key outputs:

Process Performance: D-lignin mass flow rate.
Economic Performance: Levelized Cost of Adhesive (LCOA).

Given the lack of extensive experimental or market data for this emerging technology, the authors utilized two data-generation approaches to test the inverse UQ workflow:

Single-Value Target: Using a specific target output with added Gaussian noise.
Synthetic Measurement Data: Generating multiple synthetic observations to mimic experimental variability.

Key Results

Dimensionality Reduction: Sobol GSA identified that just 5 dominant variables (Kraft lignin mass flow rate, SPI price, Kraft lignin price, and activation energies for BCD reactions 1 and 2) captured over 99% of the total output variance. This reduced the input space from 17 independent dimensions to 5.
Uncertainty Reduction:
- In the single-value target approach, applying VIBES reduced predictive uncertainty bounds by 86.3% for LCOA and 85.5% for D-lignin flow rate.
- In the synthetic data approach, uncertainty reductions were even higher, reaching 90.6% for LCOA and 93.9% for D-lignin flow rate.
- The paper notes that reducing the space to only the top variable ( $N_{top}=1$ ) resulted in high apparent uncertainty reduction but failed to adequately represent the target data, highlighting the necessity of selecting an appropriate reduced dimension.
Computational Efficiency: The reduced-space approach significantly lowered computational costs. For the 5-dimensional case, VI optimization time was reduced by approximately 31.3% (single-value) and 34.9% (synthetic data) compared to performing inference over the full 17-dimensional space.
Surrogate Accuracy: The neural network surrogate used for VI training achieved low Root Mean Squared Error (RMSE) values (~1.15% for LCOA and ~0.26% for flow rate), validating its use for gradient-based optimization.

Significance and Claims
The paper claims that VIBES provides a generic, scalable pathway for integrating global sensitivity analysis, Bayesian estimation, and posterior predictive UQ with rigorous process simulation. Its primary significance lies in:

Scalability: It enables Bayesian UQ for high-dimensional, complex chemical process systems where traditional sampling methods are prohibitive.
Fidelity: By using a surrogate only for the optimization step and propagating uncertainty through the high-fidelity model for the final results, it balances computational speed with accuracy.
Decision-Making: The framework supports uncertainty-aware design-space exploration and decision-making for biomass valorization technologies by quantifying how kinetic, design, and economic uncertainties interact to affect performance and cost.
Generality: While demonstrated on a bioadhesive process, the authors assert the framework is readily generalizable to other biomass conversion pathways and complex chemical process systems.

The authors conclude that while multiple measurements provide more informative inverse estimation than single-value targets, the core value of VIBES is its ability to systematically identify dominant uncertainty sources and reduce predictive uncertainty by more than 80% across all model outputs, even when only a reduced set of inputs is optimized.