This perspective examines how machine learning can explore molecules and materials while reducing resource use across data generation, training, inference and laboratory workflows. It estimates computational carbon footprints from reported resource inputs and proposes reuse of open datasets and models, physically informed representations, active learning, multi-fidelity calculations and model distillation. The central contribution is a resource-aware discovery framework; its benefits depend on the task and infrastructure, and reliable transfer to real conditions and experimental synthesis rema

Chemical discovery · Digital Discovery, 2026

AI can help us find useful molecules and materials. The sustainable move is to decide which questions deserve expensive calculations—and reuse what we already know.

An interactive reading of a perspective by Leonardo Medrano Sandonas and colleagues

What if progress meant learning more from the calculations we already have?

01 / Rethink the discovery loop

Which calculations earn their place?

“Chemical space” is the enormous set of possible molecules and materials. Explore a proposed way to navigate it: reuse knowledge, seek informative examples, then test the shortlist.

Your virtual experiment

Give the search a strategy.

Every point is a possible arrangement of atoms. Copper routes send a candidate for a detailed calculation.

Start with Select informative cases. Watch where the routes go.

It resolves how electrons shape atomic interactions. These calculations supply reference examples for faster machine-learning predictions.

A sign that a model has less reliable knowledge in a region. It helps choose useful new examples; physically meaningful indicators also matter when electron behavior is complex.

Based on §§2.3.2, 4.1.1, 4.3.1 and Figure 7.

Sending every candidate for a calculation repeats work in familiar regions. The paper proposes making the whole discovery workflow more selective.

Familiar chemistry: reuse a suitable model, then validate that its predictions work for this task.

The aim is scientific value per unit of computation. Selecting examples also takes work. The perspective argues for assessing the complete loop—from generating data to training, prediction and real-world validation.

02 / Account for the whole pipeline

The model is only one part of the bill.

A fast prediction tool needs examples to learn from. Building those datasets can be a much larger computational investment than training one model. Sharing and reuse make that investment more useful.

Where the carbon footprint sits

Carbon dioxide equivalent (CO₂e) expresses greenhouse gas emissions as the amount of carbon dioxide with an equivalent warming effect. Logarithmic scale: each grid interval is ×10. One tonne (t) is 1,000 kg.

OMat24 data generation is estimated at 1,250 tonnes CO₂e; training GRACE-1L-OMAT at 73.67 kg CO₂e.

The authors use Green Algorithms to estimate operational emissions from reported computing resources. CPU-hours and GPU-hours describe different hardware and are not directly comparable. Training estimates omit other stages of the pipeline and are not full lifetime emissions.

OMat24: data generation estimate; GRACE-1L-OMAT: training estimate. These are selected entries from Table 1, not matched lifetime assessments.

Source: Table 1 and Figure 1, pp. 3–4 ↗

A dataset is a community investment. Its value grows when others can reuse it, adapt existing models and avoid creating the same examples again.

03 / Look beyond the algorithm

Same work. Different electricity.

How a computer gets its electricity also shapes the footprint. The supplementary calculator files provide two country-average scenarios for the same MAD-1.5 computation. Change the electricity mix and watch the estimate change.

Keep the computation. Change the supply.

Both saved scenarios use the same energy and facility overhead. Only the electricity’s carbon intensity changes.

Carbon intensity means the greenhouse gases associated with generating one unit of electricity.

Energy: 304,180 kWh in both files. Facility overhead, also called PUE: 1.25 in both.

Germany: 333.77 g CO₂e/kWh . Switzerland: 40.76 g CO₂e/kWh .

These are the electricity-mix values saved in the authors’ country scenarios, not current grid readings. They are separate from the paper’s Table 1 headline estimate for MAD-1.5.

Germany CSV ↓ · Switzerland CSV ↓

Germany scenario: 101.5 tonnes CO₂e, from the same 304,180 kWh of energy.

Saved calculator estimates · rounded to one decimal tonne

04 / Bring the loop together

Smaller steps. Better questions.

The perspective connects sustainable computation to how discoveries are proposed, shared and tested. Open each idea to see its role in the loop.

Start from what is known.

Pretrained models can be adapted with smaller new datasets. Symmetry-aware models also respect physical relationships: rotate the atoms, and predicted forces should rotate consistently.

Let a teacher train a specialist.

An accurate but slow “teacher” model supplies examples to fit a smaller, faster “student” suited to a particular task. The tradeoff should be assessed across the workflow.

Ask for a useful material.

Inverse design starts with a desired property and proposes candidates. A useful search also considers elemental abundance, toxicity, recyclability and whether a material can be synthesized.

Let experiments teach the model.

Self-driving labs automate experimental tests and feed results back into candidate selection. The next choice balances a promising result against learning something new.

Continue with the original research

A framework for the whole journey.

This is a perspective: a synthesis of existing approaches and a proposed direction for sustainable discovery. The interactive landscape illustrates those ideas; the footprint panels use the authors’ reported-resource estimates and saved calculator scenarios.

The sustainable discovery workflow

Carbon estimates across datasets and models

Paper, authors and supplementary inputs

Towards sustainable exploration of chemical spaces with machine learning. Digital Discovery (2026). Published 16 September 2026. 10.1039/D6DD00157B.

Supplementary calculator files: lmedranos/susML. Reproduced inputs are distributed under the MIT License, copyright © 2026 Leonardo Medrano.

The graphical abstract is an AI-generated conceptual illustration based on Figures 1 and 7. It depicts the workflow concept, rather than an experimental apparatus.