Fine-tuned language models outperform larger general-purpose AI on fuel property prediction

Researchers at Tsinghua University have shown that small, fine-tuned open-source language models can outperform a much larger general-purpose model at predicting fuel physicochemical properties directly from molecular structure.

The team has named the framework FuelProp-LM. The work, from the Department of Energy and Power Engineering and the Center for Combustion Energy, led by Professor Bin Yang, was published in the journal ENGINEERING Energy.

Find your next fuel analyser in our international directory.

Why fuel property prediction needs a new approach

Predicting properties such as ignitability, sooting tendency, volatility and thermodynamic behaviour is central to fuel design and engine development.

Conventional Quantitative Structure–Property Relationship (QSPR) modelling typically requires a separate model, and manually selected molecular descriptors, for each property.

Incorporating new measurements into an existing QSPR model can also require retraining, limiting how easily such models scale or update.

A single model, one input string

FuelProp-LM instead uses a molecule's Simplified Molecular-Input Line-Entry System (SMILES) string – a standard text representation of chemical structure – as its sole input.

It predicts multiple fuel properties from a single, unified model.

The researchers fine-tuned four compact open-source language models, each under ten billion parameters, on 100,000 data entries sourced from PubChem.

No task-specific fuel property data was used at the fine-tuning stage, which the researchers say allowed the models to learn transferable structural reasoning rather than memorising property values.

Retrieval instead of retraining

To improve prediction accuracy at inference time, the team built a dynamic in-context retrieval system using MACCS molecular fingerprints and Tanimoto similarity search, run via the Hierarchical Navigable Small World (HNSW) algorithm.

This automatically retrieves structurally similar reference molecules from updated property databases to build informative prompts for each prediction. New reference data can be incorporated this way without retraining the underlying model.

Outperforming a far larger model

Across ten fuel property datasets spanning ignitability, sooting tendency, volatility and thermodynamics, the fine-tuned model consistently matched or outperformed the general-purpose DeepSeek-V3.2 model.

This held despite DeepSeek-V3.2's substantially larger parameter count, and the fine-tuned model also compared favourably against established machine learning baselines built specifically for QSPR tasks.

The researchers examined why instruction tuning improved accuracy so markedly, using an attention-layer analysis of the fine-tuned model's intermediate Transformer layers.

Fine-tuning shifted the model's attention toward the property description and the most structurally similar reference example.

It also strengthened attention to the target molecule's own SMILES string, and reduced attention paid to repetitive or less relevant prompt text.

"FuelProp-LM bypasses the tedious process of manual descriptor construction and task-specific model retraining," said Yang.

"By offering high accuracy, data scalability and deployment ease, FuelProp-LM provides an accessible and flexible tool for future AI-assisted sustainable fuel design."

What this could mean for fuel testing

For laboratories and process analysers working across fuel and petrochemical testing, the approach points toward a different operating model for property prediction tools.

Rather than maintaining a library of separate QSPR models, each tied to a specific descriptor set and property, a single fine-tuned language model updated through retrieval could in principle be extended.

This could cover new fuel formulations, including biofuels and synthetic fuels, as new reference data becomes available.

Whether that scales beyond the ten properties tested here, and how the approach performs against proprietary industrial QSPR tools rather than open-source baselines, remains to be demonstrated.

Read the full paper here.