PaperlayerView exact revision on arXiv

Source-linked AI summary

Machine learning-guided synthesis of advanced inorganic materials

Bijun Tang, Yuhao Lu, Jiadong Zhou, Han Wang, Prafful Golani, Manzhang Xu, Quan Xu, Cuntai Guan, Zheng Liu

arXiv:1905.03938v1cond-mat.mtrl-scics.LGphysics.comp-ph

TL;DR

Material synthesis is challenging, motivating strategies to optimize and accelerate it. The paper applies ML to CVD-grown MoS2 and hydrothermal CQDs, using predictions and an active feedback loop; the CVD model achieved AUROC 0.96, and the hydrothermal model improved CQD PLQY.

Abstract

from arXiv · show

Synthesis of advanced inorganic materials with minimum number of trials is of paramount importance towards the acceleration of inorganic materials development. The enormous complexity involved in existing multi-variable synthesis methods leads to high uncertainty, numerous trials and exorbitant cost. Recently, machine learning (ML) has demonstrated tremendous potential for material research. Here, we report the application of ML to optimize and accelerate material synthesis process in two representative multi-variable systems. A classification ML model on chemical vapor deposition-grown MoS2 is established, capable of optimizing the synthesis conditions to achieve higher success rate. While a regression model is constructed on the hydrothermal-synthesized carbon quantum dots, to enhance the process-related properties such as the photoluminescence quantum yield. Progressive adaptive model is further developed, aiming to involve ML at the beginning stage of new material synthesis. Optimization of the experimental outcome with minimized number of trials can be achieved with the effective feedback loops. This work serves as proof of concept revealing the feasibility and remarkable capability of ML to facilitate the synthesis of inorganic materials, and opens up a new window for accelerating material development.

photoluminescence quantum yield. Progressive adaptive model is further developed,

Multi-variable material synthesis is unpredictable, costly, and often explored through laborious trial and error, while synthesis remains less studied in ML-guided materials research. The paper applies ML to CVD-grown MoS2 and hydrothermal carbon quantum dots, and introduces an adaptive feedback approach for reducing trials.

photoluminescence quantum yield. Progressive adaptive model is further developed,

  • CVD synthesis involves many variables, making conditions unpredictable and optimal settings costly and time-consuming to find by trial and error.The text also notes that these challenges apply to hydrothermal, CVT, ALD, and MBE methods.Source: photoluminescence quantum yield.… · p. 2
  • Material synthesis is less studied in ML research than material exploration and property prediction.The paper also distinguishes prior ML-guided synthesis work focused on reactor reaction spaces and mechanisms from large-scale material synthesis.Source: photoluminescence quantum yield.… · p. 3
  • For CVD-grown MoS2, the authors construct a classifier to predict synthesis success from process parameters and recommend favorable conditions.They also introduce a progressive adaptive model to guide new-material synthesis and reduce trials while maximizing experimental success rate.Source: photoluminescence quantum yield.… · pp. 3–4
  • A regression model is applied to hydrothermal carbon quantum dots to enhance photoluminescence quantum yield, extending the approach to process-related properties.The authors frame ML-guided synthesis as a proof of concept for optimizing and accelerating inorganic-material development.Source: photoluminescence quantum yield.… · pp. 2, 4

ML-Guided CVD Synthesis with High Success Rate

This section examines machine-learning-guided optimization of CVD-grown MoS₂ synthesis, including identification of key parameters and iterative use of experimental feedback. It reports high success rates and reduced experimental time in the multivariable system.

Feature Engineering

  • After reducing 19 initial process features, the final feature set retained seven synthesis parameters, including gas flow rate, temperature, reaction time, and NaCl addition.The retained features also included sulfur distance outside the furnace, ramp time, and boat configuration.Source: ML-Guided CVD Synthesis with High… · p. 5

Model Selection

  • XGBoost-C showed the best agreement with true outcomes and most stable testing performance among the evaluated classifiers, with an AUROC of 0.96.The classifiers were evaluated using ten runs of nested cross-validation.Source: ML-Guided CVD Synthesis with High… · p. 6

Optimization of synthesis condition for higher success rate

  • SHAP ranked gas flow rate highest for predicting MoS2 growth, followed by reaction temperature and reaction time.The authors note that gas flow affects precursor transport and exposure, while temperature influences reactant vapor pressure and nucleation and grain growth.Source: ML-Guided CVD Synthesis with High… · p. 7
  • 2D MoS2 was successfully synthesized under all 10 tested conditions, substantially exceeding the original MoS2 dataset’s 61% success rate; that dataset yielded MoS2 in 183 experiments.Source: ML-Guided CVD Synthesis with High… · pp. 5, 7, 8

Acceleration with progressive adaptive model (PAM)

  • The progressive adaptive model (PAM) starts from a small initial dataset, predicts “Can grow” probabilities for remaining synthesis conditions, adds each probability with its true label to training, and repeats until all remaining probabilities are below 50.0%.Source: ML-Guided CVD Synthesis with High… · p. 8

ML-Guided CVD Synthesis with High Success Rate

  • Across 1,000 PAM trials, NC mainly clustered around a mean of 189.28 (±28.89).Source: ML-Guided CVD Synthesis with High… · p. 9
  • Success rate remained stably high at around 83.60% (±5.57); average time reduction was 36.90% (±9.63), and average true positive rate was 87.13% (±13.78).The success rate had very little variance.Source: ML-Guided CVD Synthesis with High… · p. 9
  • PAM can improve the efficiency of iterative experiments, achieving considerable time reduction despite the trade-off between time reduction and true positive rate.Source: ML-Guided CVD Synthesis with High… · p. 9
  • Using the CVD-grown MoS2 dataset, model construction, optimization, and PAM demonstrated the proposed ML methodology’s potential for navigating complex multivariable inorganic-material synthesis systems.The methodology provides a possibility of achieving high success rates and time reduction.Source: ML-Guided CVD Synthesis with High… · p. 9

ML-guided hydrothermal synthesis with enhanced targeted property

The authors use regression models to guide hydrothermal CQD synthesis toward higher photoluminescence quantum yield, then test progressive adaptive model (PAM) feedback for finding strong conditions with fewer experiments.

ML-guided hydrothermal synthesis with enhanced targeted property

  • XGBoost-R outperformed the other evaluated regression models, with R2 of 0.8402, and was selected as the best model.R2 approaching one is desirable, according to the figure caption.Source: Figure 9
  • pH was the most important feature for CQD PLQY, followed by reaction temperature and reaction time.The authors relate these rankings to CQD formation and size-dependent photoluminescence.Source: ML-guided hydrothermal synthesis with… · p. 11
  • Laboratory experiments on 11 unexplored synthesis conditions recommended by the model for their highest predicted PLQY achieved 55.5% PLQY, versus the training set’s highest PLQY of 52.8%.Source: ML-guided hydrothermal synthesis with… · p. 11
  • In a typical PAM run, true yield declined as more conditions were explored, indicating that the model identified the best synthesis conditions early in the feedback loops.The text contrasts this pattern with the more random nature of the original experimental exploration.Source: ML-guided hydrothermal synthesis with… · p. 12
  • PAM found the best condition in the confined CQD dataset within 115 experiments at 99.9% confidence, reducing time by 75.37% versus 467 trial-and-error experiments.The reported time reduction is relative to trial-and-error exploration.Source: ML-guided hydrothermal synthesis with… · pp. 12–13

Conclusion and outlook

The conclusion reports ML-guided synthesis across CVD MoS2 and hydrothermal CQDs, including an XGBoost-C AUROC of 0.96, and presents progressive adaptive feedback for early-stage guidance and fewer trials. It also identifies simplified material and feature coverage, with chemistry factors bypassed, as a boundary of current models.

Conclusion and outlook

  • AUROC of 0.96 was achieved with XGBoost-C for predicting CVD MoS2 synthesis results.The model was used to optimize CVD synthesis conditions.Source: Conclusion and outlook · p. 13
  • PAM's active feedback loop guides new-material synthesis from the beginning, aiming to enhance outcomes and minimize trials.The conclusion presents PAM as a way to improve experimental outcomes while reducing the number of trials.Source: Conclusion and outlook · p. 13
  • The approach was extended across material categories to enhance the PLQY of hydrothermal-synthesized CQDs.The conclusion presents the hydrothermal system as an application beyond CVD synthesis.Source: Conclusion and outlook · p. 13
  • Current models use one material type and few features, bypassing chemistry; broader models need features such as vapor pressure, solubility, and reactivity.The authors identify more comprehensive models covering multiple material types and chemistry-related features as needed.Source: Conclusion and outlook · pp. 13–14

Methods

The methods pair experimental synthesis protocols for MoS2 and carbon quantum dots with machine-learning procedures for prediction, evaluation, and feature attribution. XGBoost, nested cross-validation, and SHAP are used to model and interpret synthesis-related data.

Methods

  • MoS2 growth used sulfur and MoO3 precursors, a Si/SiO2 substrate, and an argon-carried furnace process with designated ramping and growth conditions.The substrate was placed polished-side down; sulfur was positioned upstream, and growth was maintained for a few minutes.Source: Methods · p. 14
  • The MoS2 synthesis temperature was ramped to the designated growth temperature and maintained for a few minutes.The supplied methods passage does not specify the designated temperature or ramping rate.Source: Methods · p. 14
  • Carbon quantum dots were synthesized hydrothermally from sulfamide solution and sodium citrate in a Teflon-lined stainless-steel autoclave.The autoclave was held at 80–300 ℃ for 0.1–12 h.Source: Methods · p. 14
  • The resulting CQDs were filtered and concentrated, then dialyzed against renewed ultrapure water to obtain the final S,N-CQDs.The dialysis used a 500 Da bag for two days, with water renewed every 10–12 h until Na+ was below the detection limit.Source: Methods · pp. 14–15
  • XGBoost predicts through an ensemble of decision-tree base estimators, trained by minimizing loss plus a complexity-penalizing regularization term.Its additive training strategy adds one tree at a time, choosing the tree that optimizes the objective at each step.Source: Methods · p. 15
  • Nested ten-fold cross-validation generated predicted probabilities for 300 samples; each outer fold used nine folds as the development set and recorded probabilities for the remaining 30, while inner stratified ten-fold cross-validation determined hyperparameters.Source: Methods · p. 15
  • SHAP assigns each input feature an attribution value using conditional expectations of the model output for feature subsets.The evaluation also defines true positive rate, success rate, and time reduction for the progressive adaptive model.Source: Methods · p. 16

Related papers

Loading 1905.03938v1…