In this talk, I will introduce Atinary’s AI platform and self-driving labs technology solutions – SDLabs – and will explain how users can significantly accelerate and enhance R&D with a data-driven approach. Additionally, Loïc will present various use-cases demonstrating unprecedented acceleration and target achievements.

Scientists and operators can connect to SDLabs in the cloud and seamlessly integrate lab equipment and off-the-shelf robotic platforms into their workflows after just a few hours of onboarding. This allows companies and R&D labs to deploy AI and machine learning solutions seamlessly, without requiring coding or ML expertise, starting with simulations or directly within their existing wet lab workflows, with or without robots.

Atinary’s AI solutions enable users to tackle complex optimization and discovery challenges that current methods cannot handle, including multi-objective and multi-parameter optimizations, categorical variables, and physicochemical descriptors. We also provide algorithms for various constrained optimizations.

The development of new energy technologies, essential for transitioning to a sustainable
future, relies on the discovery of new materials. Over the past decades, materials simulations
have significantly accelerated the discovery process, complementing experimental
approaches. These simulations offer unique insights into the fundamental mechanisms that
drive material behavior. Additionally, they can predict material properties and elucidate the
relationship between atomic structures and their properties, thereby enabling a rational design
of materials with specific characteristics. Despite their success, the discovery process has
traditionally been slow, requiring iterative cycles between theoretical predictions and
experimental verifications until optimal materials are identified, synthesized, and tested in real
devices. This paradigm has recently been broken by the creation of Materials Acceleration
Platforms (MAPs), where AI-orchestrated collaboration between AI-accelerated materials
simulations and self-driving laboratories enables closed-loop materials discovery.
In my talk, I will first discuss the development of a technology-agnostic, autonomous, and
standardized modelling framework and its integration into a MAP. The foundation of this
infrastructure is a dynamic workflow management system capable of orchestrating calculations of thermodynamic and kinetic properties, which play a fundamental role for many energy technologies. Within this framework, we have established the first autonomous workflow to discover new electrodes and solid-state electrolytes for the batteries of the future. Beyond batteries, this technology-agnostic workflow can be applied to discover new materials for a wide range of next-generation energy technologies, from fuel cells to photovoltaics. While workflows are commonly used for bulk materials, the investigation of interfaces often relies on manual, time-consuming methods based on trial and error. I will describe our efforts to implement autonomous workflows for interfaces and integrate them with the design of bulk
structures, using our work on understanding and controlling the solid/electrolyte interface in
Li-ion batteries as an example. To fully realize the potential of a MAP, a seamless data
infrastructure is required, which is capable of handling curated data and metadata from
multiple sources and with varying levels of fidelity. At the end of my talk, I will present our
approach to developing such a data infrastructure. This includes achieving complete
interoperability of computational workflows and electronic laboratory notebooks from different
sources. These involve various simulation engines, time and length scales, and automated
data collection and metadata annotation in an ontology-compliant format.

Automation in experimental battery research is limited, and cell assembly and cycling often require labor-intensive steps. Additionally, the outcomes of lab-level battery research are not always reproducible and are often dependent on the skill of the researchers. Extending lab automation improves reproducibility, accelerates experiments, and frees experimentalists from repetitive tasks, providing more time for creativity.

In a joint collaborative effort, the Swiss company Chemspeed Technologies and Empa developed and validated an automated coin cell assembly robot integrated into an argon glove box. The robot can assemble 32 coin cells per batch, with anode/cathode capacity balancing fully automated to a precision of 0.01 mg. It is capable of formulating complex mixtures of liquid electrolytes, which are then dispensed with a precision of 1 µL. Cells are then cycled on a 256-channel potentiostat interfaced with an open-source Python package developed within the Battery2030+ BIG-MAP Aurora project¹. Each cell can be traced and monitored as a digital twin within the open-source workflow management platform AiiDA, developed at EPFL/PSI². The data generated will be ontologized and made FAIR (findable, accessible, interoperable, reusable) using the BattINFO ontology, adhering to principles that facilitate data sharing and reuse.

We present the first results from robotic cell assembly and cycling, demonstrating the power of the Aurora platform in accelerating battery materials research.

The fast assessment of the global minimum adsorption energy (GMAE) between catalyst surfaces and adsorbates is crucial for large-scale catalyst screening. However, multiple adsorption sites and numerous possible adsorption configurations for each surface/adsorbate combination make it prohibitively expensive to calculate the GMAE through density functional theory (DFT). Thus, we designed a novel multi-modal transformer called AdsMT to rapidly predict the GMAE based on surface graphs and adsorbate feature vectors without any site-binding information. The AdsMT model effectively captures the intricate relationships between adsorbates and surface atoms through the cross-attention mechanism, hence avoiding the enumeration of adsorption configurations. Three diverse benchmark datasets were constructed, opening new avenues for further research on the challenging GMAE prediction task. Our AdsMT framework demonstrates excellent performance by adopting the tailored graph encoder and transfer learning, achieving mean absolute errors of 0.09, 0.14, and 0.39 eV, respectively. Beyond GMAE prediction, AdsMT's cross-attention scores showcase the interpretable potential to identify the most energetically favorable adsorption sites. Additionally, uncertainty quantification was integrated into our models to enhance the trustworthiness of the predictions. While primarily focused on heterogeneous catalyst screening, our multi-modal approach has potential applications across materials science and chemistry.

During the past decade, automated high-throughput research has evolved from “toy problems” to enabling instances of materials development and translation on compressed timelines. Self-driving labs (SDLs) take this a step further, by integrating automated high-throughput experimental (HTE) hardware with computational planning tools (inverse design algorithms, optimization algorithms, and advanced data-management systems) in closed learning loops. Although the allure of rapid progress pulls interest to SDLs, achieving tangible results requires a strategic coupled investment in both infrastructure and research. This investment necessitates a deliberate, thoughtful approach contrary to the typical rush for immediate outcomes.

This talk will provide an overview of the opportunities and many practical challenges when implementing high-throughput autonomous research systems in a university-lab setting, including: designing, troubleshooting, and de-bottlenecking high-throughput synthesis and characterization workflows, overcoming the “synthesis challenge” in its various forms, environmental control during synthesis and post-processing, managing lead safety, and the human element. I will share some modest successes to date, including discovery of new perovskite-inspired materials and optimization of existing ones, and conclude with a perspective of the opportunities ahead.

Perovskite solar cells (PSCs) have recently achieved certified efficiencies of 26.7% for single-junctions and 33.9% for silicon-perovskite tandems, drawing significant attention from academia and industry. Spin coating remains the dominant fabrication method for high-efficiency PSC devices due to its simplicity and low cost. However, manual spin coating introduces variability, such as inconsistent solution dripping speed, pipette-to-substrate distance, and timing between steps. These issues, along with variations from researcher handovers, can affect device performance.

Robotic automation offers a solution to these challenges, ensuring reproducibility and minimizing human error. One key area for automation is the anti-solvent dripping step, crucial for inducing perovskite crystallization. This process is sensitive to variables like pipette position, angle, and speed, which manual methods struggle to control precisely.

In this study, we used the fully automated KMAP system (KRICT Multi-layer Automated Spin Coating System for Perovskite Devices) to finely control anti-solvent dripping. KMAP mimics human-like experimentation with a multi-joint robot arm, capable of handling 40 samples and 16 solutions per experiment, enabling high-throughput research. 

We compared the effects of various anti-solvents (Toluene, Diethyl ether, Ethyl acetate, and Trifluorotoluene) and dripping speeds (5–25 mm/s). For low dielectric constant solvents like Toluene and Diethyl ether, slower dripping speeds yielded better results, while faster speeds worked best for high dielectric solvents like Trifluorotoluene. We also examined the impact of humidity, finding that Toluene and Diethyl ether were more sensitive to changes.

Through precise control and automation, KMAP allowed us to analyze the complex thin-film formation process, demonstrating the effectiveness of integrating robotics and AI to tackle experiments that are difficult for humans to perform manually.

Perovskite solar cells have seen significant advancements in power conversion efficiency (PCE) in recent years, reaching record PCEs above 26 %. However, achieving consistent performance across different laboratories remains a challenge due to the variability inherent in manual processing methods[1,2]. Our study addresses this challenge by demonstrating the potential of a fully automated spin-coating robot for fabricating perovskite thin films. The commercial spin-coating robot autonomously performs sample positioning, pipetting, timed anti-solvent dispensing, and annealing. Compared to manual methods, this automated system provides excellent repeatability and improved homogeneity of perovskite thin films. Transferring an established perovskite composition to this automated fabrication process enabled champion PCEs as high as 19.9 %, comparable to manually processed perovskite solar cells. Through a series of nine batches produced over two months, we evaluated the device performance and the crystallinity and optoelectronic properties of fully automated processed perovskite absorbers. Consistent peak positions and ratios in X-ray diffraction analysis confirm the repeatability of the composition and crystallographic structure. Minimal variation in photoluminescence emission peak wavelengths and implied open circuit voltage indicate the consistency of optoelectronic characteristics. Our work provides a foundation for automated systems supporting research in perovskite solar cells. It aims to accelerate developing and deploying high-performance, high-throughput, and highly repeatable fabrication processes.

The corresponding paper is submitted and under review.
D.O. Baumann, F. Laufer, J. Roger, R. Singh, M. Gholipoor, U.W. Paetzold, submitted (May 2024)

     Halide perovskites have emerged as one of the most promising and diverse material systems in the history of photovoltaics. Intense interest in these materials is a result of several favorable optoelectronic characteristics while being solution-processable using abundant elements at low temperatures. Furthermore, bandgap tunability from the UV to the NIR makes them ideally suited for multi-junction solar cells.

     Despite their advantages, most perovskite compositions used in highly efficient PSCs exhibit comparatively poor thermal stability due to the presence of organic cations, particularly the volatile methylammonium. Inorganic Cs-based perovskites are an intriguing exception and have recently been demonstrated to exhibit excellent operational stability in single-junction solar cell devices. An attractive next step would be to find a stable composition for use in multi-junction devices. The optimization problem therefore is therefore to find a stable material with a suitable bandgap that can be processed at temperatures compatible with a multi-junction substrate. 

     Here we report on a robotic high-throughput system for exploring the compositional space of perovskite top-cell absorbers. The system comprises a robotic liquid-handling system capable of preparing a large array of solutions (up to approx. 100 per batch) from a set of stocks or precursors ready for manual spin-coating. A separate apparatus has been assembled for automatic characterization and accelerated aging. In the latter, optoelectronic properties (e.g., bandgap, luminescence) are recorded before and during ageing under elevated temperatures and intense illumination. The gathered information is used to develop a material database, and computer-aided decision-making (e.g., Bayesian optimization) is used to model the hyperspace and iteratively improve the model with each new experiment. In combination we expect this system to dramatically accelerate the pace of optimization and the discovery of commercially relevant perovskite compositions for use in tandem and multi-junctions.

In this presentation we will address topics such as the comparative stability of films fabricated by spin-coating and drop-casting, the effect of selective contact layers on aging behavior, and the role of the halide fraction.