Machine-learning-assisted materials discovery using failed experiments

Abstract

Inorganic-organic hybrid materials such as organically templated metal oxides, metal-organic frameworks (MOFs) and organohalide perovskites have been studied for decades, and hydrothermal and (non-aqueous) solvothermal syntheses have produced thousands of new materials that collectively contain nearly all the metals in the periodic table. Nevertheless, the formation of these compounds is not fully understood, and development of new compounds relies primarily on exploratory syntheses. Simulation- and data-driven approaches (promoted by efforts such as the Materials Genome Initiative) provide an alternative to experimental trial-and-error. Three major strategies are: simulation-based predictions of physical properties (for example, charge mobility, photovoltaic properties, gas adsorption capacity or lithium-ion intercalation) to identify promising target candidates for synthetic efforts; determination of the structure-property relationship from large bodies of experimental data, enabled by integration with high-throughput synthesis and measurement tools; and clustering on the basis of similar crystallographic structure (for example, zeolite structure classification or gas adsorption properties). Here we demonstrate an alternative approach that uses machine-learning algorithms trained on reaction data to predict reaction outcomes for the crystallization of templated vanadium selenites. We used information on 'dark' reactions--failed or unsuccessful hydrothermal syntheses--collected from archived laboratory notebooks from our laboratory, and added physicochemical property descriptions to the raw notebook information using cheminformatics techniques. We used the resulting data to train a machine-learning model to predict reaction success. When carrying out hydrothermal synthesis experiments using previously untested, commercially available organic building blocks, our machine-learning model outperformed traditional human strategies, and successfully predicted conditions for new organically templated inorganic product formation with a success rate of 89 per cent. Inverting the machine-learning model reveals new hypotheses regarding the conditions for successful product formation.

Auditing black-box models for indirect influence

“…The Dark Reactions data shows different top ranked features for the three models, though all three rankings include the minimum Pauling electronegativity and the maximum Pearson electronegativity in the top ranked cluster of features. These values are calculated for the inorganic components of the reaction and have been shown to be important for distinguishing between chemical systems in this data set [20], which this audit confirms. Features indicating the presence of elements and amounts of metal elements with specific valence counts similarly allow the models to classify chemical systems.…”

“…The SVM and FNN top features include atomic properties of the inorganics that are related to the electronegativity of an element, so these proxies are correctly also highly ranked. The top ranked decision tree features additionally include the average molecular polarizability for the organic components, which was previously hypothesized as important to synthesis of templated vanadium selenites explored via this data set [20]. For all three models, the lowest ranked descriptors are constants scored correctly as having no influence to the model.…”

“…Machine-learning models now determine and control an increasing number of real-world decisions, from sentencing guidelines and parole hearings [2] to predicting the outcome of chemical experiments [20]. These models, powerful as they are, tend to also be opaque.…”

Computational development of the nanoporous materials genome

“…if we are given a structure, can we predict its performance, say in CO 2 separations, based on its surface area, pore size, or pore volume? There have been several recent studies dedicated to answering this question using models developed by machine learning 32,45,114,[131][132][133][134] , a field that is becoming extremely powerful in materials science 135,136 . It was shown that 1-dimensional geometric descriptors are able to successfully predict adsorption at high pressures 134,137 and low temperatures 114 , using representative datasets of materials to train machine learning models.…”

Human versus Robots in the Discovery and Crystallization of Gigantic Polyoxometalates

“…[5] Fore xample,r ecently,W icker and Cooper [6] applied machine learning methods to draw am ap of crystallinity according to the size of am olecule and its number of rotatable bonds.S imilarly,O liynyk et al [7] used machine learning to predict structures of inorganic binary compounds of the general formula AB by considering various atomic and physical properties in their calculations.O f particular interest, Norquist et al [8] made use of data from unsuccessful syntheses to predict reaction outcomes of vanadium compounds and compared the efficiencyo ft heir algorithms with the typical strategies that human chemists apply.…”