AURORA - An Automatic Robotic Platform for Materials Discovery (2025)

The discovery and design of novel high-performance materials for specific applications present a formidable challenge due to the vast chemical space that encompasses numerous potential candidates. This complexity is amplified by the considerable time and resources required to systematically screen and evaluate all possible combinations. For example, metal-halide perovskites (MHPs), currently the most promising photovoltaic materials, adhere to a general formula ABX₃ referring to their 3D family of compounds, where A is a monovalent cation that can be either organic, such as CH₃NH₃⁺ (MA⁺) and HC(NH₂)₂⁺ (FA⁺), or inorganic, like Cs⁺. B is typically a divalent p-block metal cation (e.g., Pb²⁺ and Sn²⁺), forming an inorganic framework of BX₆ octahedra linked together via halide anion ligands, including I^–, Br^–, or Cl^–. (1) In addition, low-dimensional perovskite-like materials (forming 2D, 1D, and 0D structures) can be synthesized when other ions are introduced to the sites that cause the 3D network of BX₆ octahedra to collapse into lower dimensions of sheets, chains, or even isolated complexes. (2−4) Another strategy to expand the family of materials related to the 3D perovskites is to replace two divalent metal cations in the B position with one monovalent metal ion and one trivalent metal ion, resulting in a new group of double-perovskite-type materials with the formula A₂B^IB^IIIX₆. (5) Already these small compositional modifications lead to a countless number of members in the library of MHP-based materials, not to mention the consequences of further expansion by doping or alloying those pure MHP-like compositions.

Screening a selected subgroup of materials can represent a huge challenge, characterized by repetitive and painstaking experimental work. For these reasons, an automated screening platform is highly desired and beneficial to reduce routine human labor and minimize errors thereby unlocking a straightforward strategy for systematic studies of material properties. (6,7) At present, robotic platforms have been introduced to different areas of materials discovery, such as a robotic platform for the synthesis of colloidal nanocrystals, (8) a mobile robot-based platform focusing on organic synthesis, (9) an automated biomateriomics platform, (10) and the Poseidon robotic platform for the discovery of lithium-ion battery electrolytes. (11) There have also been several platforms developed specifically for MHPs and perovskite-related technologies. (12−14) In this work, we introduce our new modular robotic system, AURORA, as a new member of the robotic platforms family. The system is designed as an automatic robotic platform for the screening of functional materials, which includes robotic synthesis, materials characterization, and evaluation with respect to a specific device application. MHPs have been selected as an illustrative case study. By combinatorial synthesis of mixed halide perovskite polycrystals, followed by photoluminescence investigation, the automatic fabrication of arrays of mesoscopic perovskite solar cells (mPSCs), and dynamic current–voltage (IV) performance characterization, we demonstrate the capability of the AURORA system for the screening of perovskite-like materials. The future development of the AURORA system into a self-driving platform for data collection, reflecting on the balance between quantity and quality in building reliable data sets for training new machine learning (ML) models, is also discussed.

The framework of the AURORA system is shown in Figure 1a. It was designed to perform synthesis, characterization, and application with ideas retrieved from existing systems in the literature, databases, and theoretically predicted results. The automated experiments provide possibilities to generate process-controlled data of high quality, which can offer feedback not only to the materials genome for future and more accurate predictions but also to the robotic system itself for self-optimization.

The AURORA system, used in this work for the MHPs case studies, consists of three fully integrated subunits: a robotic synthesis unit, a device test module, and a flexible robot arm. The robotic synthesis unit includes a liquid handling robot equipped with high-precision liquid handling labware (pipettes, tubes, and well-plates), a controllable temperature module, and a cooling stage. The robot arm can move samples between different units. In addition, there is a semiintegrated plate reader as a characterization module for photoluminescence spectroscopy (PL). An illustrative overview of the platform is given in Figure 1b.

The example MHP materials selected in this study are evaluated in mesoscopic solar cells. Such cells are sometimes referred to as monolithic solar cells, where mesoporous materials are precombined into a solar cell substrate, only missing the active photovoltaic material that can be simply added in the form of a precursor solution percolating through the top porous materials with subsequent solvent evaporation. The substrates used were based on a configuration of FTO/cTiO₂/mTiO₂/mZrO₂/mCarbon; the prefix c relates to a compact layer and m to a mesoporous layer. (15) This printable mesoscopic structure has proven useful in laboratory research of MHPs due to comparable conversion efficiencies with respect to the conventional type of PSCs, high stability, and unique advantages in terms of cost and scalability for industrialization. (16) Moreover, this configuration allows solution deposition and target material crystallization in the preprinted inorganic scaffold to constitute the very last but crucial step of cell fabrication, which is highly compatible with the employment of a dispensing robot. In addition, mesoscopic cells are less dependent on the excellent film-forming properties of the new materials generated, which may be a limiting factor in the discovery of new materials. As demonstrated by previous work from our group, the mesoscopic PSCs have shown excellent compatibility for photovoltaic materials screening. (17) In this work, we have designed a new printable mesoscopic substrate array (Figure S1), allowing the simultaneous screening of 16 devices on a single substrate, with the potential to further upscaling to 32 or 64 cells on a single substrate and the implementation of parallel analysis.

The robotic solar cell device test module was designed for the 16-cell substrate architecture. A custom-made multichannel potentiostat assembly (P&L Scientific AB) was designed to consist of a single-channel potentiostat acting as a master unit, a power supply, a substrate holder containing 16 apertures with a mask area of 0.126 cm², and three printed circuit boards (PCBs): a multiplexer board, a light-emitting diode (LED) board, and a board with spring-loaded contacts. The 16 LEDs were calibrated (details are given in Figure S2) to 1000 lx, which corresponds to a power intensity of 250.8 μW cm^–2, (18) which is a commonly used light intensity when studying the application of a material in indoor photovoltaics. IV characterization of the device performance involves the key parameters: short-circuit photocurrent density (J_sc), open-circuit voltage (V_oc), fill factor (FF), and power conversion efficiency (PCE), which were extracted via a control Python script. It is worth noting that the performance of the devices under dim light not only highlights the potential of the studied materials for indoor applications but also provides a reasonable correlation with their performance under 1 Sun (AM 1.5G, 100 mW cm^–2) illumination. Example comparisons are given in Figure S3.

A proof-of-concept workflow and the corresponding units of the robotic platform are illustrated in Figure 2. The combinatorial synthesis was carried out in a liquid-handling unit. The resulting solutions were dispensed either into a 96-well plate for photoluminescence investigation or onto a preprinted mesoscopic inorganic scaffold array substrate of 16 solar cells. An antisolvent was added to the well plate for the precipitation of polycrystalline materials for evaluation. Following precipitation, the well plate is transported to the plate reader for characterization. In this work, we demonstrate the successful application of a workflow aiming at the synthesis of mixed halide perovskite materials with a new combination of solvent–antisolvent as compared to previous studies (Note S1). (19,20) The PL results of mixed methylammonium lead triiodide (MAPbI₃) in γ-valeroacetone (GVL) and methylammonium lead tribromide (MAPbBr₃) in dimethylformamide (DMF) using acetic acid (AcOH) as the antisolvent are shown in Figure S4.

After precursor solution deposition, the solar cell substrates were transferred by the robot arm to the temperature module for solvent evaporation and crystallization. After programmed heating and cooling procedures, the robot arm transported the substrate to the cell test module for IV scans. The substrate can be programmed to be transferred between the temperature module and the IV-test module several times to repeat the heating-measurement cycle for dynamic monitoring of the device performance during the crystallization process.

This first example experiment is based on the well-known MAPbI₃, one of the prototype MHPs. The fully automated workflow can be found in Note S2 and Video S1. Specifically, 16 solar cells were fabricated with 2 μL of MAPbI₃ precursor solution dispensed on top of each cell. The heating temperature was set to 90 °C for 2 min, while the cooling temperature was 30 °C, with an additional 60 s on the cooling block. The heating–cooling characterization cycle was iterated 6 times, with two IV characterizations per cycle. Figure S5 shows the resulting 12 IV curves of the 16 solar cells. All 16 cells show similar performance with small deviations, indicating the reliability of the data acquired from the robotic platform. The variance of the performance parameters between the different cycles demonstrates the necessity of the iterative procedure applied, which enables dynamic monitoring of the performance parameters of the target materials. This dynamic evaluation allows for the identification of the optimal conditions for high performance, especially when there are differently screened components with unknown optimal crystallization conditions. This stage is also highly suitable for initial machine-learning implementation into the robotic system.

From the IV-curves and extracted data, the reverse scan curves in some systems show nonideal shapes, leading to nonphysical, overestimated maximal power points, and fill factors. This phenomenon has frequently been reported for PSCs with carbon-based counter electrode materials, although the underlying reasons remain ambiguous. (21−23) Therefore, the key parameters from forward scans were extracted for further discussion. As shown in Figure 3, there are obvious differences in the first 6 scans, i.e., three preparation cycles, especially regarding the increase in PCE, V_oc, and J_sc. This indicates that the conditions selected, involving a minimum of 3 preparation cycles, lead to a satisfactory degree of crystallization of the target material in the mesoscopic inorganic scaffolds. After 3 preparation cycles, the PCE and V_oc only increase marginally upon the introduction of more preparation cycles, while J_sc instead tends to decrease. The FFs of all investigated samples remain stable at around 0.4, with a gradual improvement as more preparatory cycles are applied. The relatively low FFs observed represent one trade-off required when an automated approach is used, which will be further discussed below. Overall, the standard deviation in the performance parameters for the different solar cells is within an acceptable range for screening purposes.

For comparison, one batch with 16 mesoscopic PSCs was manually fabricated using the same conditions as the robotized workflow, including the volume of solution, the heating temperature, set times, and cycles. However, the pipetting of precursor solutions and transfer of the substrates were carried out manually. The extracted parameters from the IV scans are shown in Figure S6. In general, the manually produced data show a trend similar to that of the robot-fabricated cells. However, the subtle variance in the pipetting and transfer by hand leads to different performance. Here, the high robustness and reproducibility of the robotized approach emerge as an additional and intentional benefit.

Mixed halide perovskite materials were also investigated in the robotic platform. Different compositions of the MAPbX₃ (X = I, Br) precursor solutions were dispensed onto the 16 mesoscopic solar cell array. The percentage of iodide studied ranged from 30% to 100% in steps of 10% iodide (labeled as I3 to I10), and 2 cells were prepared for each composition. The heating–cooling characterization parameters were set to 90 °C for 5 min heating and 30 °C with an additional 60 s on the cooling stage, iterated for 6 cycles with 2 IV scans for each cycle. The resulting IV curves are shown in Figure S7, and the extracted forward PCE versus the cycle of iteration is shown in Figure 4. It is worth mentioning that, in our system, we focus on the heating–cooling iteration procedures and not a fixed heating time. As a consequence, 2 cycles of 5-min heating do not necessarily equal 5 cycles of 2-min heating. From the figures obtained, it can be deduced that the iodide-rich samples generally display higher PCEs than the bromide-rich samples under the selected cycling conditions. Overall, the MAPb(I_0.7Br_0.3)₃ composition shows balanced performance and stability during the time of investigation, offering a highest PCE of 23.0% in cycle 3 and 21.8% in cycle 6, which may correspond to around 5% under 1 sun illumination according to the comparison in Figure S3. For better comparison and further improvement of the multicomponent MHPs, an adjustment of the workflow─especially regarding the heating–cooling-testing parameters─will be included in future studies.

In addition to the direct synthesis and subsequent analysis, the AURORA platform is also designed for post-treatments of the synthesized materials and devices, enabling dynamic analysis of the samples that are exposed to physical stress, chemical treatment, or simple on-the-shelf storage. In this study, the samples are transferred between storage and characterization modules. More complex post-treatments, such as heat treatment, light exposure, and solution treatment, will be included in future work. Figure S8 shows the resulting time-dependent PL spectra for the 11 synthesized MAPb(I_xBr_1–x)₃ (x = 1 to 0) mixtures. The extracted peak wavelength and peak intensity are mapped against storage time to visualize the dynamic changes in the peaks during storage (Figure 5a). Generally, there is better stability in the bromide-rich samples, indicated by their smaller changes in peak wavelengths and intensities, which can be explained by the stronger interaction between Pb²⁺ and Br^–. (24) The storage investigation was also carried out for the robot-fabricated MAPbI₃ devices. As shown in Figure 5b, all performance parameters improve after storage at ambient conditions (20 °C, ∼40% RH) for 1 day. Longer storage times under ambient conditions lead to degradation, as observed by a decrease in J_sc. However, V_oc and FF continue to gradually increase, leading to a relatively stable PCE of the cells stored. This is consistent with published reports, (25) and the observed behavior is explained by the competition between moisture-induced further crystallization and decomposition of the materials.

When designing a materials screening platform, there are inherent trade-offs, including throughput vs accuracy, speed vs depth, data quantity vs quality, versatility vs specialization, and automation vs flexibility. For example, the well-plate-based combinatorial synthesis followed by optical characterization has shown great potential for use in high-throughput experiments and for providing reliable optical properties of the screened materials. However, the automated procedure is not as accurate as a detailed manual study, which includes well-separated steps such as synthesis combined with structural analysis involving X-ray crystallography and electron microscopy, in terms of ensuring product structure and phase purity. The insights obtained from the static PL-spectroscopic evaluation are also limited compared to time-resolved techniques. This limitation also applies to the design of the device system in this work, including the use of mesoscopic solar cell arrays and the LED-based light source. It should be emphasized that our automated approach does not focus on the fabrication of a device with all technical parameters optimized but rather provides a fast way to understand the composition-structure-performance correlations of a target materials subgroup in devices such as solar cells.

For mesoscopic solar cells, very slow crystallization of the absorber in the inorganic scaffold benefits the device performance. In some reports, a cover is applied during the heating step to slow down the evaporation of the solvent used and, thus, the crystallization process. (26) We tested this by manually fabricating 16 cells with MAPbI₃ and heating them at 70 °C for 30 min using a cover (Figure S9). The average PCE from forward scans was (13.08 ± 3.36)%, with the best cell showing a reverse PCE of 22.6%; significantly better than a MAPbI₃ solar cell heated on an “open” hot plate fabricated by the robot. However, this approach does not allow dynamic PCE monitoring and usually requires a significantly longer preparation time for better crystallization, which is not optimal for an automated process.

The versatility and flexibility should also be considered for a robotic platform. In the AURORA system, the materials characterization is currently more versatile than the device module, which has been designed for a specific type of solar cell configuration and measurement, although both still have limitations with respect to the scope of materials that can be studied. A high degree of automation is efficient for predefined workflows and provides more controlled and transferable parameters for reproducibility, avoiding human errors. However, it may also make the system less adaptable and less flexible. For this reason, we have designed the platform with reorganizable modules, which allow adaptation of the automation level according to the screening scenario and facilitate the implementation of new functional modules for materials or device characterization step-by-step. This also allows AURORA or each module used in this work to be easily integrated with other robotic materials platforms. It is notable that the core robotic modules offer significant capabilities with respect to cost, and the flexible expansion by implementation of new modules for characterization is facilitated by the large number of the current availability of relatively inexpensive table-top instruments.

In order to provide an overview of the current state of AURORA, we present a comparative table (Table S1) that highlights its existing capabilities, strengths, and areas for future improvement. For comparison, we have selected some robotic platforms that focus on MHPs, considering the model materials used in this work, although there are many other systems available, and these robotic platforms, including ours, may also be applicable to other types of materials.

In order to further improve AURORA and balance the aforementioned trade-offs of a screening platform, the next steps in robot extension will be to add modular powder X-ray diffraction and diffuse reflectance spectroscopy units. This approach is highly compatible with a hierarchical screening workflow and the use of hybrid screening methods. For example, a workflow can include prescreening of perovskite-inspired materials, starting with the solution-based synthesis in the current liquid-handling module and high-throughput photoluminescence analysis in the plate-reader module, followed by moderate-throughput phase identification for the selected promising candidates. Ultimately, the most promising materials can be studied in the device module. Modules for battery evaluation are currently under construction to expand the screening scope of materials relevant to renewable energy projects, such as electrolyte materials for Li-ion batteries. (11) In addition, as mentioned before, the robotic platform is highly suitable for machine-learning implementation. For instance, a Bayesian optimization process integrated into the synthesis-characterization workflow can acquire the optimized device performance with minimal experimental trials for interesting materials/device candidates through a real-time feedback loop. It is also possible in the future to implement inverse design to identify optimal compositions for new materials based on the collected materials and/or device screening data. In order to achieve this, it is essential to further enhance the data reliability and standardize the workflow through the robotic platform.

In summary, we have developed the AURORA robotic screening platform for materials discovery, demonstrated through initial experiments with the well-known MAPbX₃ materials family (X = Br, I). Our proof-of-concept experiments highlight the capability of AURORA for perovskite materials synthesis, PL spectroscopic characterization, and dynamic performance evaluation, including the development of a novel mesoscopic solar cell array. The automation protocols reduce human effort and error in the data collection, enabling the creation of a reliable, high-quality database suitable for self-learning algorithms and facile data transfer to other users. Although still in an early phase of development, the AURORA system operates through an automated workflow that efficiently screens materials subfamilies. Looking ahead, by integrating feedback from current experiments and adding more modules of characterization, we envision expanding the system’s capabilities to explore a broader range of materials, unlocking new possibilities and applications. This will facilitate the evolution of AURORA into a robust, self-driving platform, ready to meet the dynamic challenges of modern materials research and discovery.