Autonomous materials synthesis by machine learning and robotics Open Access

Materials are a powerhouse of innovation,¹ and accelerating the development of new materials is vital for a sustainable society. Currently, new materials development in laboratories involves repeated cycles of conception, synthesis, and characterization, manually performed by researchers. The inclusion of machine learning, robotics, and big data into these cycles promises to revolutionize materials research and beyond.

To begin, let us sketch the concept of a next-generation materials research laboratory. Future lab equipment and instruments should be Connected, Autonomous, Shared, and operate in a High-throughput manner (CASH, Fig. 1). With the ongoing evolution of machine learning and robotics, the CASH movement is steadily spreading around the world and is expected to have a transformative effect on basic research.

If CASH was implemented in a materials laboratory, then new materials with the desired properties could be rapidly synthesized using a closed-loop approach.^2,3 Researchers would first prepare the raw material, let the machine-learning-driven apparatus and robotics synthesize and evaluate the material properties, and publish the results. Here, publishing would involve much more than writing a manuscript; all results—from the synthesis conditions (explanatory variables) to the physical properties (response variables)—would be automatically collected into a central database and shared. In addition, any information collected during the synthesis would be accumulated in a central database. In essence, future CASH laboratories would become data production factories.

To implement CASH, we first consider the materials development and optimization process as the following stepwise process: (1) determination of the synthesis conditions, (2) synthesis, (3) evaluation, and (4) determination of the next synthesis conditions (Fig. 1). Traditionally, each task in this cycle is manually performed by researchers. Even when the process is automated, the determination of the next conditions requires the judgment of researchers. This determination is a critical step that reduces throughput. The development of machine learning and robotics promises to make the autonomous closed-loop cycle possible, removing all researcher intervention after the start of the process, at least for routine work.

While the closed-loop concepts are beginning to be applied to biosystems^4–6 and organic synthetic chemistry,^7,8 few examples focus on solid-state materials such as polymers or inorganic materials.⁹ Focusing on inorganic materials, storing synthesis conditions in a database is critical. For example, TiO₂ can be either a semiconductor or an insulator depending on specific synthesis conditions. Thus, with the ever-increasing demand for complex materials and exotic properties,¹⁰ the materials database should incorporate synthesis condition-dependent properties.

In this paper, we highlight our recent experiments that serve as examples of the CASE approach and demonstrate its great promise in materials discovery in materials science. We first describe the experimental design necessary for CASH implementation, including the apparatus based on machine learning and robotics. We then demonstrate the autonomous deposition of a TiO₂ thin film with a targeted physical property using the apparatus. Finally, we discuss the significance of the inclusion of machine learning, robotics, and big data into future materials research and present our vision on how to integrate human knowledge, experience, and intuition.

We have designed and custom-built a closed-loop system that combines Bayesian optimization,^11–14 automatic synthesis, and automatic physical property evaluation [Fig. 2(a)]. The system consists of a robotic arm located at the center of a hexagonal chamber, which is attached to six satellite chambers equipped with an automatic film deposition apparatus and automatic evaluation system [Fig. 2(b)]. The robotic arm performs all sample transfers between the satellite chambers. The deposition conditions needed to obtain optimal physical properties are predicted using Bayesian optimization. The predicted conditions are fed-back to the automatic film deposition apparatus, achieving full autonomous closed-loop cycles.

To demonstrate the CASH approach, we aimed to minimize the electrical resistance of anatase Nb-doped TiO₂ thin films.¹⁵ The minimization is achieved by varying the oxygen partial pressure (Po₂) during the deposition of the TiO₂ thin film. The closed-loop cycles proceed as follows. Before starting, a stock of glass substrates is loaded into the load-lock chamber. This is the only manual step of the process; every step hereafter is automated. Next, the robot arm transfers a substrate to a film deposition chamber, where a TiO₂ thin film is deposited under the synthesis conditions indicated by Bayesian optimization. After the deposition, the film is transferred to the physical-property evaluation chamber, where a probe electrode is pressed against the film for resistance evaluation. The evaluated results are sent to the computer, which adds the latest Po₂ and the electrical resistance values to the previous data. The system then uses Bayesian optimization to predict the Po₂ that leads to minimal electrical resistance. The learning continues by repeating this cycle until convergence is reached, i.e., a thin film with minimal electrical resistance is obtained. We demonstrate the closed-loop optimization for one variable (one-dimensional experiment).

We used reactive magnetron sputter deposition, a method that bombards the TiO₂ target with high-energy Ar gas to scatter TiO₂ for deposition onto a glass substrate. Two sets of experiments were performed: one with a Ti_0.94Nb_0.06O₂ target (T1) and the other with a Ti_1.98Nb_0.02O₃ target (T2); the significant difference between T1 and T2 is the Ti:O ratio, with T2 being a more reduced condition (both targets have a purity of 99.9% and 2 in. diameter, Toshima Manufacturing Co., Ltd.). The Nb doping was used to increase TiO₂ conductivity.¹⁶ The value of Po₂ was controlled by changing the flow ratio of two gases: one was pure Ar gas and the other was a gas mixture of Ar (99 mol. %) and O₂ (1 mol. %). The range of oxygen partial pressures was set to 2 × 10⁻⁴ Pa–5 × 10⁻³ Pa, considering the results of earlier experiments.^17,18 The total gas flow rate and gas pressure were fixed at 10 SCCM and 0.50 Pa throughout the deposition, respectively. The radio-frequency power supply at the target was maintained at 100 W. The typical film thickness is ∼120 nm with a growth time of 1 h. For the Ti_0.94Nb_0.06O₂ (T1) target, the thickness of the film deposited at the lowest P_O2 (2 × 10⁻⁴ Pa) and at the highest P_O2 (5 × 10⁻³ Pa) was ∼120 nm and ∼100 nm, respectively. Because the two thickness values are close, the difference in the thickness does not affect the minimization of electrical resistances, which change by orders of magnitude depending on P_O2. The substrate temperature was maintained at room temperature during the deposition. After deposition, the samples were annealed in a high vacuum with a background pressure of ∼10⁻⁵ Pa at 450 °C for 30 min to crystallize the film.

An automatic system was custom-built for evaluating the electrical resistance of the thin films. After film deposition, the sample was allowed to cool for about 20 min to room temperature. Then, the sample was transferred to this system, where a motor-driven electrode probe came into contact with the sample to evaluate its electrical resistance.

Bayesian optimization is a powerful tool for finding the global maximum/minimum of black-box functions with a small number of trials. In this work, we used a standard Bayesian optimization algorithm with Gaussian process regression for the prediction of the black-box function.

The main task of the Bayesian optimization is to predict the Po₂ that produces thin films with minimal electrical resistance. This prediction is based on Gaussian process regression. To initiate Bayesian optimization, we first collected electrical resistance data for films deposited under preset Po₂ conditions, namely, the smallest and the largest values and middle of the P_O2 range. The optimization process was then performed in cycles. Each cycle consists of three steps: (1) selecting a value of Po₂ for the next deposition using the data obtained so far, (2) depositing the thin film under that P_O2, and (3) evaluating electrical resistance, followed by addition of that data to the dataset. Steps (1)–(3) are iterated repeatedly, and optimization is completed when convergence is achieved.

In this system, human intervention is only needed to decide whether the cycle has converged. In the future, this decision will be made by the computer as well. This system enables new methods for producing thin films with desired physical properties within a reasonable amount of time. It can also be used for multi-dimensional optimization involving several film formation parameters (oxygen partial pressure, substrate temperature, annealing temperature, nitrogen partial pressure, film formation rate, etc.).

We used the Python package GPy for Gaussian process regression. The grid for P_O2 was linearly divided into 128 grids in the range from 2.0 × 10⁻⁴ Pa to 5.0 × 10⁻³ Pa. The radial-basis function was used as a kernel, where the variance and the length scale were set to 0.3 and 6 × 10⁻⁴ (3 grids), respectively. The lower confidence bound was adopted as the acquisition function, where its hyperparameter was set to 5.0. and the preset three conditions were sparsely selected, at both ends (2.0 × 10⁻⁴ Pa and 5.0 × 10⁻³ Pa) and at the center (2.7 × 10⁻³ Pa) of the grid. These conditions were set manually in advance. The fourth and subsequent deposition conditions were determined by Bayesian optimization sequentially, and previously used conditions were not selected again.

Figures 3(a)–3(c) illustrate the closed-loop cycles performed to minimize the T1 thin film resistance. As shown with the red dots, the Bayesian optimization algorithm first searches across a wide range of P_O2 [Fig. 3(a)] before converging toward the optimal Po₂ = 7.29 × 10⁻⁴ Pa [Fig. 3(b)], reaching it at the fourteenth cycle [Fig. 3(c)]. After the fourteenth cycle, the Bayesian optimization selected P_O2 stayed near the optimal value, suggesting the convergence. The trend of P_O2 vs electrical resistance from previous experiments indicates that there is only one minimum within the searching space.^17,18 Therefore, we conclude that the obtained minimum is the global minimum.

Similarly, the closed-loop cycles for minimizing the T2 thin film resistance are illustrated by Figs. 3(d)–3(f). Here, the Bayesian optimization converged toward the optimum P_O2 = 2.43 × 10⁻³ Pa, reaching it at the eighteenth cycle [Fig. 3(f)]. Compared to T1, the entire curve for T2 is shifted to the right to a more oxidizing condition, as is consistent with the lower oxygen content of T2.¹⁹ 

The experiments on T1 and T2 yielded two positive results. The first is technical: the use of Bayesian optimization identified the global resistance minimum from several local minima caused by the reduced (e.g., Ti₂O₃) and Magneli phases [Ti_nO_2n−1 (n = 3–9)].

The second is temporal: the use of the autonomous closed-loop cycle has improved the experiment throughput by an order of magnitude. As the time stamps in Table I show, the system is capable of fabricating thin films at a rate of ∼2 h per sample. Thus, ideally, ∼12 samples can be fabricated within a day when the closed-loop cycles are operated continuously. The present system took more time to complete the convergence because the substrates had to be resupplied by the researchers. Manually, researchers can perform two film depositions per day on average, taking seven days to complete the fourteen-cycle optimization. Adding three days to account for the weekend and other work- and lifestyle-related essentials, ten days is a reasonable estimate. Accordingly, throughput has increased by tenfold.

In this study, we used a constant length scale of 3 grids to optimize an experimentally well-known relationship between electrical resistance and P_O2. In general, to find the global maximum (or minimum) in a small number of trials, it is critical to tune the length scale. For this purpose, one may use a relatively larger length scale at the initial stage and a smaller length scale as the number of observation data is increased.

The actual magnitude of the improvements discussed above will only be evident when the CASH approach is applied to novel materials development. Unlike the T1 and T2 experiments, which only considered P_O2 to minimize resistance, discoveries of novel materials with exotic properties will need to consider other dimensions such as materials composition, deposition temperature, sputtering pressure, and power input to the target. Manual optimization of such multidimensional systems will involve too many sets of manual experiments to be feasible. The CASH approach will use Bayesian optimization to consider every dimension at once and robotics to fabricate and evaluate samples, making the synthesis of new materials achievable in a short time. This approach will usher in the next paradigm shift in materials research and beyond.

Materials science is headed toward a data-driven future. In this evolution, machine learning, robotics, and big data are turning a previously inconceivable idea into a real possibility: the understanding of materials in their entirety. As discussed above, traditional research usually synthesized and characterized a new material for a few exotic properties and ignored other properties for lack of resources. As illustrated in Table II, databases of substances and properties remain incomplete. The completion of the database will not only uncover hidden properties, but it will also lead to the discovery of new rules that determine material properties through the analysis of correlations and trends. Filling these gaps to create a database is one of the important objectives of materials science.

To fill the gaps, the autonomous closed-loop system is a powerful tool. Currently, we are expanding the instrument in Fig. 2 into a multiunit system that can automatically characterize a list of properties such as electron transport, optical, dielectric, and magnetic properties (Fig. 4). With this new system, once a material is synthesized, many physical properties will be collected automatically to construct a database (Table II). This database will contain reliable non-best data useful for improving the performance of machine learning.

The construction of such databases enables us to discover trends that would not emerge with current research practices. Here, we have demonstrated two experiments with ten times the human throughput. However, the true revolution will occur when individual instruments are expanded into multiunit systems, and data from all multiunit systems are shared as big data to be processed by advanced machine learning (materials informatics).^20–22 

The example experiments presented here use the simplicity of thin film materials to demonstrate how machine learning and robotics can be applied for materials science. In practice, further advances in machine learning and robotics are required. It is clear that machine learning will play a key role in the formulation of future scientific theories. By including technologies such as explainable machine learning²³ and symbolic regression,²⁴ big data can be reduced down to simple, human-understandable expressions. This reduction would stimulate researchers to provide physical insights into materials and discover new laws of physics (human-in-the-loop). For robotics, a teaching technology to learn the motion of experimentalists is critical to incorporate researcher’s experience. Taken together, breakthrough occurs when researchers of machine learning, robotics, and materials collaborate to inspire materials science.

Finally, we note that machine learning and robotics will not work without expert materials scientists asking the right questions from the start. To this end, the training of future materials scientists must evolve; they need to understand what machine learning can solve and set the problem appropriately. With good problem setting, machine learning and robotics will significantly accelerate materials research. It is vital to educate future materials scientists with the abilities of problem setting and foresight to take on this challenge.

In this paper, we used our recent experiments in autonomous synthesis and resistance minimization of thin films as an example to illustrate how the speed and the volume of future big-data collection in materials science will be achieved. The approach that combines machine learning, robotics, and big data demonstrates the tremendous potential in materials science. It is only through coevolution with such technologies that future researchers can work on more creative research, leading to the acceleration of materials science research.

See the supplementary material for schematic videos of our closed-loop system in Fig. 2 (closed_loop_movie.mp4) and the expanded system designed to evaluate many physical properties in Fig. 4 (multiunit_system_movie.mp4) and animations of the closed-loop cycles for both Ti_0.94Nb_0.06O₂ thin films in Figs. 3(a)–3(c) (TiO2_movie.mp4) and Ti_0.99Nb_0.01O₂ thin films in Figs. 3(d)–3(f) (Ti2O3_movie.mp4).

The data that support the findings of this study are available within the article (and its supplementary material).

This research was supported by JST-CREST (Grant No. JPMJCR1523) and JST-PRESTO (Grant No. JPMJPR17N6). We thank Dr. Daniel Packwood for the critical reading of this manuscript. We also thank Dr. Patrick Han from SayEdit.com for manuscript editing.