🤖 machine learning

Prediction of Diffusion Coefficients in Mixtures with Tensor Completion

This paper presents a hybrid tensor completion method enhanced by Bayesian priors and active learning strategies to accurately predict temperature-dependent diffusion coefficients in binary mixtures, overcoming the limitations of existing single-temperature matrix completion models and semi-empirical approaches.

Original authors: Zeno Romero, Kerstin Münnemann, Hans Hasse, Fabian Jirasek

Published 2026-07-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Zeno Romero, Kerstin Münnemann, Hans Hasse, Fabian Jirasek

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

The Big Picture: Filling in the Blanks

Imagine you have a giant spreadsheet (a matrix) where every row represents a specific chemical "solute" (like a drop of dye) and every column represents a specific "solvent" (like water or oil). Inside each box of the spreadsheet, you want to know how fast that dye spreads through that oil. This speed is called the diffusion coefficient.

The problem is that this spreadsheet is mostly empty. Scientists have only measured a tiny fraction of these combinations because doing the experiments is slow, expensive, and tedious.

For a long time, scientists used "semi-empirical models" (like the SEGWE model) to guess the missing numbers. Think of this like using a rough map based on old rules of thumb. It works okay, but it's not very precise.

Recently, scientists started using Machine Learning (ML) to fill in the blanks. They treated the problem like a game of "Guess the Missing Number" (called Matrix Completion). If you know how a specific dye behaves in water, and how a different dye behaves in oil, the computer can mathematically guess how that first dye behaves in oil. This worked great, but only for one specific temperature at a time (usually room temperature, 298 K).

The New Innovation: The 3D Puzzle

The authors of this paper asked: "What if we need to know the speed at 300 K, 310 K, or 350 K? Can't we just run the computer again for each temperature?"

The answer is no, because there isn't enough data for those other temperatures to teach the computer.

So, they built a Tensor Completion Method (TCM).

The Analogy: If the old method was a flat 2D spreadsheet (Rows x Columns), the new method is a 3D Rubik's Cube.
- Dimension 1: The Solute (the dye).
- Dimension 2: The Solvent (the oil).
- Dimension 3: The Temperature.

Instead of training three separate computers for three separate temperatures, they trained one single 3D computer model. This model looks at the whole cube at once. It learns that "heat makes things move faster" and uses that pattern to fill in the empty spots not just for the temperatures they have data for, but for any temperature in between.

How They Made It Smarter: The "Hybrid" Approach

The computer is smart, but it's also a bit of a blank slate. To help it, the authors used a Hybrid approach:

The "Teacher" (SEGWE): They fed the computer predictions from the old, rough SEGWE model first. This gave the computer a basic "physics lesson" on how things should behave.
The "Student" (The AI): Then, they let the computer look at the few real experimental data points they actually had. The computer adjusted its "teacher's" advice to match reality.

This is like giving a student a textbook (the SEGWE model) and then having them take a few practice tests (the real data). The student learns the general rules but corrects them based on what actually happens in the real world.

The "Active Learning" Strategy: Asking the Right Questions

Even with a smart model, they still needed more real data to make it perfect. But they didn't want to waste time measuring random chemicals.

They used a strategy called Active Learning.

The Analogy: Imagine a teacher grading a test. Instead of asking the student to solve 100 random math problems, the teacher looks at the student's answers, sees where they are confused (high uncertainty), and says, "Let's just practice these three specific problems you keep getting wrong."

The computer looked at its own predictions, found the chemical mixtures it was least sure about, and told the scientists: "Go measure these specific ones." The scientists then went into the lab and used a special machine (PFG NMR) to measure exactly those mixtures. This "targeted" approach improved the model's accuracy much faster than just guessing randomly.

The Results

The paper claims that this new 3D "Rubik's Cube" model (TCM) is better than the old 2D spreadsheet models and the old rough maps (SEGWE) in three ways:

Accuracy: It predicts the diffusion speeds more correctly.
Range: It can predict speeds at temperatures where they have no data at all (extrapolating between 268 K and 378 K), whereas the old models could only guess at the specific temperatures they were trained on.
Robustness: It makes fewer wild, incorrect guesses (outliers) than the other methods.

Summary

In short, the authors took a flat, temperature-limited machine learning model and turned it into a 3D, temperature-flexible model. They taught it using a mix of old physics rules and new real-world data, and they used a "smart question" strategy to get the most helpful new experiments done. The result is a much better tool for predicting how chemicals mix and move at different temperatures.

Technical Summary: Prediction of Diffusion Coefficients in Mixtures with Tensor Completion

Problem Statement
Predicting diffusion coefficients in liquid mixtures is critical for numerous industrial and natural processes, yet experimental data remains scarce due to the vast number of possible mixtures and the labor-intensive nature of diffusion measurements. While machine learning (ML) offers alternatives to established semi-empirical models, existing Matrix Completion Methods (MCMs) are limited to single-temperature predictions. Their accuracy is heavily dependent on the availability of high-quality experimental data for each specific temperature, and they cannot extrapolate to temperatures outside their training set. Furthermore, data availability is often sparse at temperatures other than the standard 298 K, hindering the development of reliable models for broader industrial operating ranges.

Methodology
The authors propose a hybrid Tensor Completion Method (TCM) designed to predict temperature-dependent diffusion coefficients at infinite dilution ( $D^\infty_{ij}$ ) in binary mixtures. The methodology integrates the following components:

Tensor Representation: Unlike MCMs which treat data as a 2D matrix (solute $\times$ solvent), the TCM structures the data as a third-order tensor with dimensions for solutes, solvents, and temperature.
Tucker Decomposition: The model employs a Tucker decomposition to factorize the tensor. It utilizes latent feature vectors for solutes ( $u$ ) and solvents ( $v$ ), which are temperature-independent, and a temperature-dependent feature vector ( $w$ ). These are combined via a core tensor ( $\kappa$ ) to reconstruct the diffusion coefficients.
Hybrid Bayesian Training: The TCM is trained using a two-step Bayesian framework:
1. Pre-training: The model is fitted to a fully populated synthetic tensor generated by the semi-empirical Stokes-Einstein–Gierer–Wirtz (SEGWE) model. This incorporates prior physical knowledge.
2. Fine-tuning: The posterior distributions from the pre-training step serve as informed priors for a second training step on the sparse experimental data tensor. This allows the model to adapt to real-world data while retaining physical constraints.
Active Learning (AL): To address data scarcity, the authors employ an AL workflow. An uncertainty sampling strategy selects solute-solvent pairs with the highest prediction uncertainty for new experimental measurement via Pulsed-Field Gradient (PFG) NMR. A specific rule was introduced to prevent redundancy by removing compounds from the sampling pool after they have been sampled consecutively too many times.
Data Scope: The initial training data consisted of experimental $D^\infty_{ij}$ values at 298 K, 313 K, and 333 K. The model was subsequently evaluated for its ability to extrapolate linearly to temperatures between 268 K and 378 K.

Key Contributions

Development of a Hybrid TCM: The paper introduces a novel method that extends matrix completion to tensor completion, enabling the prediction of diffusion coefficients across a continuous temperature range rather than at fixed points.
Integration of Physical Priors: By embedding SEGWE model predictions as priors within a Bayesian framework, the method improves accuracy even when experimental data is sparse.
Active Learning Integration: The study demonstrates a workflow where ML-driven uncertainty sampling guides the acquisition of new experimental data (via PFG NMR), specifically targeting 19 new solute-solvent systems to enhance model performance.
Extrapolation Capability: The TCM successfully extrapolates to temperatures not present in the training set (268 K–378 K), a capability lacking in standard MCMs.

Results

Predictive Accuracy: The hybrid TCM outperformed both the established SEGWE model and the isothermal MCMs across all studied temperatures (298 K, 313 K, and 333 K). The TCM exhibited lower Absolute Relative Errors (ARE) and narrower interquartile ranges, indicating greater robustness and fewer outliers.
Impact of Active Learning: Incorporating the 19 newly measured diffusion coefficients (at 298 K, 313 K, and 333 K) into the training set yielded a substantial improvement in the TCM's predictive accuracy.
Generalization: The model demonstrated the ability to generalize to continuous temperature values, providing reliable predictions in the range of 268 K to 378 K, including temperatures where no training data existed.

Significance
The authors claim that this work highlights the potential of combining data-efficient machine learning methods with adaptive experimentation. By addressing the limitations of single-temperature models and the scarcity of experimental data, the proposed hybrid TCM offers a more reliable tool for predicting transport properties. This advancement is significant for applications requiring diffusion data across wide temperature ranges, where experimental measurement is often impractical. The study validates that integrating physical priors (SEGWE) with ML and actively acquiring data can significantly advance predictive modeling in thermodynamics.