Predictive Inorganic Synthesis based on Machine Learning using Small Data sets: a case study of Hydrodynamic Diameter-controlled Cu Nanoparticles

arXiv:2512.16545 · cond-mat.mtrl-sci, cs.LG · Submitted 2025-12-18 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Predictive Inorganic Synthesis based on Machine Learning using Small Data sets".

Jane: The gist: Ensemble regression models successfully predict hydrodynamic diameters with good predictive performance given the limited dataset, and an ensemble model achieved an R2 of 0.74,

Tom: First, who's behind it and why it matters.

Paper summary: Jane: So looking at the whole paper "Predictive Inorganic Synthesis based on Machine Learning using Small Data sets: a case study of Hydrodynamic Diameter-controlled Cu Nanoparticles", what's the big picture here for anyone who isn't deep into materials science?

Tom: The main point is that you don't need massive datasets to use machine learning effectively when you are working with small, high-quality data sets, and they proved that ensemble regression models can successfully predict hydrodynamic diameters for copper nanoparticles with an R2 of zero point seven four and an MAE of twenty-three point eight one nm

Lu: What this means is that you can accelerate synthesis optimization because traditional one-variable-at-a-time approaches are slow, but ML accounts for those multidimensional correlations in the synthesis outcomes thirteen fourteen <ref:2512.16545#pg2,traditional one-variable-at-a-time>

Meng: From a deployment standpoint, the fact that they developed a model that retains all experimentally relevant parameters makes it more suitable as a synthesis guidance tool for someone trying to get specific results

Jane: It really shows that even in chemistry where every experiment takes time and resources to create data points, you can still use these data-driven methods to guide the process toward your desired outcome

Tom: And they showed that understanding the distribution shape, whether it's mono-modal or multi-modal, is also important because that classification helps you know when your prediction for the hydrodynamic diameter is actually reliable

Lalam: The study concludes by showing that this workflow is intended for interpolation within the experimentally investigated synthesis parameter space

Jane: So in simple terms, they are giving researchers a way to use ML to guide them in synthesizing these copper nanoparticles with better control over their size and shape than just guessing parameters

Tom: That's what it boils down to. They're using ML on small data to build a practical tool for controlling nanoparticle synthesis outcomes

Conclusion: Tom: So, we've been diving into this paper about predicting copper nanoparticle size using machine learning on small datasets.

Jane: It looks like the title is "Predictive Inorganic Synthesis based on Machine Learning using Small Data sets: a case study of Hydrodynamic Diameter-controlled Cu Nanoparticles." That sounds a little technical for someone who isn't in materials science, right?

Lu: It's about using machine learning to predict how big copper nanoparticles will be, specifically focusing on their hydrodynamic diameter. The key here is that they are working with what they call small datasets.

Meng: Small datasets are always a headache for deploying models at scale. So, if the model works well on this small set, does that mean it actually helps us make more copper nanoparticles reliably?

Tom: It means the model showed good predictive performance. They got an R squared of zero point seventy-four and a mean absolute error of twenty-three point eighty-one nanometers when they focused on mono-modal size distributions.

Jane: That's a solid number to start with. So, what does that actually change for someone in the lab? What's the practical application here?

Lalam: It means this model can act as a guidance tool. The authors used it to generate candidate synthesis parameter sets for three different target hydrodynamic diameters, and two of those candidates landed within the experimental uncertainty of their target values.

Lu: And they made sure the final ensemble model kept all the important experimental parameters in it, so it's designed to work with real synthesis conditions, not just theoretical ones.

Tom: So it's not just a prediction tool; it's something that helps narrow down what you should try next in your synthesis process.

Jane: And they also introduced a classification task—predicting if the resulting particles will be large or small, and whether the size distribution is mono-modal or multi-modal. That adds another layer of certainty to the prediction.

Meng: I see that classification part matters for knowing when you can actually trust your quantitative hydrodynamic diameter predictions. It’s about knowing what conditions are valid.

Tom: Exactly. The paper shows that even with limited data, if you use a robust ensemble regression model, you can get useful guidance for making Cu NPs.

Jane: This whole thing is about moving from just running experiments to having a smart system that suggests the best experiment next based on what it already knows.

Lu: It highlights that for lab-scale work, complex AI models might not always be the answer, but when curated with good data, they offer real assistance in navigating that parameter space.

Brent Motmans, Digvijay Ghogare, Thijs G.I. van Wijk, Joren Van Herck, Saba Heidarian, Pieter De Meyer, Berend Smit, An Hardy, Danny E.P. Vanpoucke

Institute for Materials Research (IUMAT) · Hybrid Materials Design (HyMaD) · imec Institute of Molecular Sciences and Technology (IMeC) · Energyville Institute for Materials Research (IUMAT) · Design and Synthesis of Inorganic Materials (DESINe) Institute · Electrochemistry Excellence Centre (ELEC) Flemish Institute for Technological Research (VITO), Laboratory of Molecular Simulation (LSMO) Institut des Sciences et Ingénier Chimiques Ecole Polytechnique Férale de Lausanne EPFL

cond-mat.mtrl-sci, cs.LG

Submitted: 2025-12-18

Updated: 2026-10-08

Comments: 15 pages (+16 pages SI), 5 figures (+16 SI), 4 tables (+11 SI)

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 66/100

The gist: The gist: Ensemble regression models successfully predict hydrodynamic diameters with good predictive performance given the limited dataset, and an ensemble model achieved an R2 of 0.74, an MAE of

Key concepts

Ensemble Regression ML: This technique trains multiple base models on different subsets of the data and combines their predictions. The final model is built by 'pasting' these instances together to create a more stable and accurate prediction for the hydrodynamic diameter, which is crucial when experimental data is scarce.
Latin Hypercube Sampling: This method is used to design an efficient set of experimental conditions across a multi-dimensional parameter space. It ensures that the chosen synthesis parameters (like temperature and concentration) are spread out evenly across all possible ranges, maximizing the information gained from a small number of experiments.
Hydrodynamic Diameter (DLS): This is a measurement obtained through Dynamic Light Scattering (DLS) that describes the size of particles in a liquid suspension. The paper focuses on predicting this specific physical property—the hydrodynamic diameter—using synthesis parameters like temperature and precursor concentration.
Mono-modal vs. Multi-modal Distribution: This refers to the shape of the particle size data. Mono-modal means the particles cluster around a single, distinct size. Multi-modal means there are several distinct peaks in the size distribution, indicating a mixture of different sized particles resulting from the synthesis process.

Terminology

Accuracy — 1× in this paper · explained in 1 episode
Large Language Models (LLMs) — 1× in this paper · explained in 8 episodes
While Machine Learning — 1× in this paper

Summary

The gist: Ensemble regression models successfully predict hydrodynamic diameters with good predictive performance given the limited dataset, and an ensemble model achieved an R2 of 0.74, an MAE of 23.81 nm, compared to an R2 of 0.60, an MAE of 33.54 nm for the DoE model

Data Set Construction and Experimental Setup

Latin Hypercube Sampling is used to efficiently cover the parameter space while creating the experimental data set. The investigated parameter ranges were selected based on previous Cu polyol synthesis studies and practical considerations, including a temperature range of 175 to 200 °C, a Cu precursor concentration limited to 1–10 mM, and reaction times of 1 to 20 minutes. The Cu NPs are synthesized via the microwave (MW)-assisted polyol route.

Modeling Approaches

Two modeling approaches are applied to the dataset: ensemble regression ML implemented in the AMADEUS framework, and classical DoE. Ensemble regression models use pasting-type models consisting of 100 base instances trained on different randomly selected 80/20 splits of the complete data set. For quantitative regression, only data set C is considered because it provides an unambiguous hydrodynamic diameter for every experiment, focusing on mono-modal distributions.

Model Refinement and Feature Engineering

The ensemble regression models undergo iterative refinement to reduce model size and improve generalization. The fourth generation model retains six features: [Cu], T, time, [Cu]2, [Cu]3, e[Cu], [Cu]time, and ln([Cu] × time). This final ensemble model achieved an MAE of 23.81 nm and an R2 of 0.74 when restricted to mono-modal DLS distributions.

Classification Tasks

Two classification tasks were introduced: predicting whether the resulting particles will be large or small based on synthesis parameters, and predicting whether the hydrodynamic diameter distribution is mono- or multi-modal. The classification models using both random forests and Large Language Models (LLMs) exhibited only modest predictive performance, indicating that this small dataset is insufficient to fully exploit the capabilities of complex LLMs. The SVM classifier for modality classification achieved an accuracy of 0.92 and a κ = 0.82.

Synthesis Guidance and Applicability

The ML model was used to generate candidate synthesis parameter sets for three target hydrodynamic diameters, where two yielded hydrodynamic diameters within the experimental uncertainty of their respective target values. The final ensemble ML model retained all experimentally relevant synthesis parameters within the model, making it more suitable as a synthesis-guidance tool. The classification task provides a necessary complement to the regression model by identifying conditions under which quantitative hydrodynamic diameter predictions are valid.

Conclusion on Methodology

The study demonstrates that carefully curated small data sets, paired with robust classical ML, can effectively support the synthesis of Cu NPs and highlights that for lab-scale studies, complex models like LLMs may offer limited benefits. The results further show that the synthesis outcome is not only characterized by hydrodynamic diameter, but also by the modality of the resulting DLS-derived hydrodynamic diameter distribution. The proposed workflow is intended for interpolation within the experimentally investigated synthesis parameter space.

Author Contributions and Support

Brent Motmans performed Data curation, Formal analysis, Investigation, Software, Writing–original draft, Writing–review & editing. The work was financially supported by the Special Research Fund (BOF) via the BOF project via the BOF project. The computational resources and services used in this work were provided by the VSC (Flemish Supercomputer Center), funded by the Research Foundation Flanders (FWO) and the Flemish Government–department WEWIS.

Data Availability

All data supporting this article are included in the manuscript and Supplementary Information. The following files are available free of charge • Supplementary information.pdf: Experimental synthesis parameters and datasets; DLS, UV-Vis, and SAED characterization data; ML model development and validation; DoE, SVM, and LLM analyses.

References

[1] M.-C. Daniel and D. Astruc, Gold nanoparticles: Assembly, supramolecular chemistry, quantum-size-related properties, and applications toward biology, catalysis, and nanotechnology.

[2] T. Zhai et al., Recent Developments in One-Dimensional Inorganic Nanostructures for Photodetectors.

[3] M. B. Gawande et al., Nanomaterials: a review of synthesis methods, properties, recent progress, and challenges.

[4] Q. A. Pankhurst et al., Applications of magnetic nanoparticles in biomedicine.

[5] N. Baig et al., Nanomaterials: a review of synthesis methods, properties, recent progress, and challenges.

[6] Z. J. Baum et al., Artificial Intelligence in Chemistry: Current Trends and Future Directions.

[7] K. T. Butler et al., Machine learning for molecular and materials science.

[8] J. Schmidt et al., Recent advances and applications of machine learning in solid-state materials science.

[9] B. Sanchez-Lengeling and A. Aspuru-Guzik, Inverse molecular design using machine learning: Generative models for matter engineering.

[10] G. Huang et al., Application of Machine Learning in Material Synthesis and Property Prediction.

[11] A. Agrawal and A. Choudhary, Perspective: Materials informatics and big data: Realization of the “fourth paradigm” of science in materials science.

[12] D. E. P. Vanpoucke et al., A machine learning approach for the design of hyperbranched polymeric dispersing agents based on aliphatic polyesters for radiationcurable inks.

[13] E. M. Williamson and R. L. Brutchey, Using DataDriven Learning to Predict and Control the Outcomes of Inorganic Materials Synthesis.

[14] N. A. C. Lah, Properties optimisation of nanostructures via machine learning: Progress and perspective.

[15] H. Huo et al., Machine-learning rationalization and prediction of solid-state synthesis conditions.

[16] A. Merchant et al., Scaling deep learning for materials discovery.

[17] S. Curtarolo et al., The high-throughput highway to computational materials design, Nature Mater 12, 191.

[18] C. Chen et al., A Critical Review of Machine Learning of Energy Materials, Adv. Energy Mater. 10, 1903242.

[19] P. Raccuglia et al., Machine-learning-assisted materials discovery using failed experiments, Nature 533, 73.

[20] Y. Ma et al., The mastery of details in the workflow of materials machine learning, npj Comput Mater 10, 141.

[21] A. A. Guda et al., Machine Learning Analysis of Reaction Parameters in UV-Mediated Synthesis of Gold Nanoparticles.

[22] H. Tao et al., Self-Driving Platform for Metal Nanoparticle Synthesis: Combining Microfluidics and Machine Learning, Adv. Funct. Mater. 31, 2106725.

[23] R. W. Epps et al., Artificial Chemist: An Autonomous Quantum Dot Synthesis Bot, Adv. Mater 32, 2001626.

[24] F. Pellegrino et al., Machine learning approach for elucidating and predicting the role of synthesis parameters on the shape and size of TiO2 nanoparticles, Sci Rep 10, 18910.

[25] E. J. Braham et al., Machine LearningDirected Navigation of Synthetic Design Space: A Statistical Learning Approach to Controlling the Synthesis of Perovskite Halide Nanoplatelets in the QuantumConfined Regime, Chemistry of Materials 31, 3281.

[26] D. F. Williams et al., Optimizing silver nanowire synthesis: machine learning improves and predicts yield for a polyol, millifluidic flow reactor, Appl Nanosci 13, 6539.

[27] X. Guo et al., Copper Nanoparticles on Graphene Support: An Efficient Photocatalyst for Coupling of Nitroaromatics in Visible Light, Angew. Chem. Int. Ed. 53, 1973.

[28] R. Reske et al., Particle Size Effects in the Catalytic Electroreduction of CO2 on Cu Nanoparticles, J. Am. Chem. Soc.

Improvements for AI systems

Improve predictive accuracy for quantitative size estimation by using an ensemble regression model restricted to mono-modal DLS distributions, as this reflects the use of a welldefined regression target. This will allow the system to provide a quantitative DLS-derived hydrodynamic diameter for synthesis guidance, achieving an MAE of 23.81 nm compared to DoE models.
Enhance synthesis guidance capability by implementing feature selection based on dominance importance analysis, as this process allows us to determine the influence of different synthesis parameters on particle properties. This will result in a reduced model size and improved generalization, moving from a 19-feature model to a more robust fourth generation ensemble model.
Develop a domain-aware feature engineering pipeline by incorporating nonpolynomial features like ln([Cu] × time) and [time]/T into the regression models. This allows the system to capture complex physical interactions, such as the interplay between time and temperature in a single scalar, potentially leading to better prediction accuracy.
Establish a reliable applicability domain screening mechanism using a binary classification model (SVM) that predicts the DLS distribution modality (mono- or multi-modal). This classification serves as a screening step, enabling the identification of conditions under which the regression model is expected to yield reliable quantitative predictions.
Integrate interpretable models for synthesis guidance by utilizing an explicit analytical equation derived from the fourth generation ensemble model. This provides a simple, interpretable model expressed as an explicit analytical equation further limits computational overhead during model evaluation.