Transfer-Aided Deep Learning for Life Cycle Assessment Prediction of Precursor Molecules

To date the environmental impacts of production of most chemicals are unknown, hindering sustainable process evaluation and design. Life cycle assessment (LCA) is the most comprehensive method for estimating these environmental impacts but existing LCA databases cover a small fraction of the industrial chemicals. Environmental impact prediction models promise to bridge this gap; several models were published, with the apparent limitations of limited accuracy and applicability. The very small number of available…

By the Truvace deskOctober 15, 20262 min read1 sourcePeer-reviewedTRV-2026-1500· cite & verify

In brief

Researchers applied transfer learning to deep learning models that predict the environmental impacts of producing chemicals. The models were first pretrained on molecular price data. They were then used on the scarce life cycle assessment datasets available for industrial chemicals.

This matters because most chemicals have no known production impacts, which hinders sustainable process evaluation and design. The approach only partly solves the data scarcity problem, and gains appeared in two of three impact categories, so accuracy and applicability remain limited.

Main points

Pretraining on price data improved accuracy for three deep learning architectures in two of three environmental impact categories.
Averaged across architectures and impacts, R 2 rose by 0.11 and mean percentage absolute error fell by 2.5%.
The most accurate carbon footprint model reached an R 2 of 0.62, similar to existing literature benchmarks.
The best model showed an overall prediction error of 22% for organic precursors of six active pharmaceutical ingredients.

The gain

Supervised pretraining on molecular price data raised the accuracy of deep learning models that predict environmental impacts of chemicals, easing the shortage of life cycle assessment data.

The rundown

Life cycle assessment is described as the most comprehensive way to estimate the environmental impacts of chemicals, but existing databases cover only a small fraction of industrial chemicals. The very small number of available datasets is identified as the main obstacle for prediction models.

The authors tested whether supervised pretraining on molecular price data could compensate for this shortage. Accuracy improved across three model architectures, though not in every impact category.

As a practical check, the strongest carbon footprint model was applied to organic precursors for six active pharmaceutical ingredients, giving a 22% overall prediction error.