Posters can be found at each of our venues throughout the conference.
Bahen Centre for Information Technology
Myhal Centre for Engineering Innovation and Entrepreneurship
Posters can be found at each of our venues throughout the conference.
This study used a self-driving lab to identify structured polymer components that efficiently absorb mechanical energy (K). We hypothesized that pairing components can yield superior K. The complexity of this optimization problem was tackled by computationally screening >12,000 components and implementing an active learning algorithm that fixed one component to test pairings. Besides identifying complex structures that are high efficiency, this research may shed light on how to screen for emergent properties of complex systems.
We have designed and are prototyping a self-driving laboratory simulation ecosystem in NVIDIA Omniverse and Isaac Sim to enable rapid design of laboratories and of the experiments they perform. Our primary focus is a set of tools that improves access to self-driving laboratory technologies and simplifies interfaces for faster learning how to use the system.
Here, we use an experiment that isolates high-affinity binding molecules from large diverse combinatorial libraries, called AS-MS, along with machine learning (ML) to guide the discovery of ultra-high affinity binding compounds. From the AS-MS data, the ML provided a “map” that enabled the accurate prediction of high-affinity binding across multiple libraries, natural and non-natural. We hope that this mapping approach will accelerate drug discovery and development.
We report a two-stage data-driven approach to the targeted synthesis of organic photoredox catalysts and the subsequent reaction optimization for metallophotocatalysis, as demonstrated for decarboxylative C(sp3)–C(sp2) cross-coupling of amino acids with aryl halides. The broad relevance of our approach to other areas is enhanced because our approach does not involve any specialized automation or robotics equipment; all of the chemistry is done by hand, and the efficiency stems from the use of Bayesian optimization-led searches.
We introduce Chemcrow, a chemistry agent that integrates 17 expert-designed tools to enhance performance in chemistry-related tasks such as organic synthesis, drug discovery, and materials design. Our evaluation, including both LLM and expert assessments, demonstrates ChemCrow's effectiveness. However, there are potential risks and misuse concerns, which we discuss. ChemCrow bridges the gap between experimental and computational chemistry, benefiting experts and non-experts while fostering scientific advancement.
Traditional ingredients in everyday products such as shampoo are getting eco-conscious consumers in a lather. Machine learning methods can be used to design better molecules & materials; however, they require experimental data to train on. Performing the experiments is typically time, resource and labour-intensive. To alleviate this, a high-throughput liquid formulation workflow using robotics and AI has been developed. This helps accelerate product development and brings us closer to making more eco-friendly products.
Drug and development of new functionalized materials requires macro-quantities on the gram scale to ensure enough material at the end to enable structure determination and confirmation of the desired target. Serial Nano-Electron Diffraction, with a new method for creating showers of nanocrystals on demand, provides the highest possible structural resolution, using only microgram to nanograms quantities – opening the door to true nanoscale chemistry with accompanying orders of magnitude speed up in the discovery pipeline.
Our approach optimizes chemistry experiments with a robot incorporating real-time feedback from a crowd of expert chemists into Bayesian Optimization. The robot uses an interactive website to enable experts to share their hypotheses. This AI-human collaboration allows for more adventurous search space exploration and faster convergence. Our experimental results show that this approach improves BO with expert hypotheses, even in cases where the expert's knowledge is inaccurate and can recover from misleading hypotheses.
For the development and wider adoption of self-driving labs (SDL), the availability of automated and sustainable manufacturing techniques that allow for accelerated and scalable synthesis of materials is a prerequisite. Conventional wet chemistry techniques are recipe-based and involve multiple steps for the production of nanoparticles and layers. Iterations for conventional synthesis alone usually takes many years, but also the wet chemistry steps hamper integration into an SDL. An alternative approach to these methods is the proposed single-step, dry synthesis and deposition of materials.
Automated recognition of molecular images using AI has recently emerged as an intriguing challenge. Here, we suggest an unsupervised learning model for molecular image recognition by converting the image to a graph representation that enables image reconstruction. By comparing the reconstructed image to the original one, the model will learn the accurate graph representation in an unsupervised manner. The permutation-invariant nature of the model is expected to enhance its accuracy compared to previous ones.
Recent advances in large language models (LLMs) provide powerful tools for predictive modeling on a variety of natural language tasks. Prior work developing Genome-scale Language Models (GenSLMs) demonstrated the potential for LLMs to predict future SARS-CoV-2 variants of concern prior to their emergence by modeling the evolutionary process. We propose to expand this framework by demonstrating how to effectively scale GenSLMs for bacterial genomes, and even further to more complex eukaryotic organisms including yeast and humans.
Analysis bonding patterns in high entropy alloys using EXAFS is a difficult and time consuming task that could greatly benefit from active learning (AL). Like many other potential applications of AL, the complexity of the EXAFS analysis problem complicates direct application of AL. We present our approach combining on-the-fly autoencoders with quantitative analysis based on modern statistical methods.
Our framework integrates unsupervised learning with chemistry classification, addressing overconfidence in supervised methods. By training class-specific unsupervised models, we calculate likeliness scores to measure sample similarity and avoid overconfident predictions on unseen data. Testing on benchmark datasets reveals comparable performance to supervised methods, with superior detection of out-of-sample samples. Our approach holds potential for reliable predictions in chemistry, particularly in domains where safe classification is vital.
Integrating evolution algorithms with Bayesian optimization, we enable efficient materials discovery in self driving labs.
An overlooked challenge of operating an autonomous chemistry lab is the high-level orchestration and planning required to efficiently execute multiple workflows (e.g., multiple syntheses) in a way that maximizes throughput. We view this as a scheduling and optimization problem. We have developed a discrete event simulator, “LASSO”, to serve as a digital twin for an automated synthesis platform and an environment within which to test scheduling algorithms.
Creating a community-driven ecosystem of open-source hardware designed for laboratory automation solution can help broaden the application space and increase the implementation of self-driving laboratories. Through the fabricable multi-tool motion platform, Jubilee, the project aims to create community-driven resources for the development and preservation of open-hardware ecosystems. These will allow a broader application space and enable a scientific maker movement for more inclusion of automation tools into research environments and educational curricula.
Our research unveils a new area of application for the Hammett equation: the prediction of binding energies of homogeneous catalysts. For such a simple and inexpensive method, the predictions are sufficiently accurate to identify ideal catalysts. We further enhance simplicity and accuracy with combining rules and delta machine learning, additions that can make a significant contribution to accelerating catalyst design.
We are surrounded by structures designed to absorb energy, such as padding in sports equipment or packaging for shipping. Efficient energy absorption can decrease the size and cost of the structures while increasing safety. By using a self-driving lab to perform >12,500 physical experiments, we discovered the most efficient structure for absorbing energy that has ever been observed. This large database of experiments also provides insight into the dichotomy of elastic vs. plastic materials.
The Fuel Cell INFormation Ontology (FCINFO) and Fuel Cell Value Chain Ontology (FCVCO) aim to enhance knowledge discovery, data integration, and collaboration in fuel cell research, development, and manufacturing. FCINFO concentrates on fuel cells, their components, materials, and interfaces, while FCVCO promotes knowledge integration, collaboration, and decision-making throughout different fuel cell technology development phases. Both ontologies offer a structured framework for representing concepts and relationships within the fuel cell domain.
We introduce AutoProtocol, a unique framework that uses advanced AI and robotic simulations to automate molecular biology experiments. This system, trained on textual data from sources like Science Exchange, PLOS, and Bio-protocol, generates intricate protocols. It takes into account biologist feedback and error corrections from digital twin simulations, minimizing human error and enhancing reproducibility. AutoProtocol is a significant step towards self-driving molecular biology laboratories, saving time and effort in experiment design and execution.
We propose a novel generative model for designing molecules with desired properties that assembles retrosynthetically prepared chemical building blocks to improve the synthesizability. The model can efficiently use dozens of thousands of building blocks including unseen blocks for molecule generation, which was not possible in previous models. We demonstrate the real-world applicability of our strategy by generating potential molecules in various material discovery domains.
We present a scheme for the acceleration of chemical space exploration based on chemical reaction networks. Combining fast, but inaccurate quantum mechanical calculations with smart selection strategies and slow, but accurate quantum chemical methods allows for constructing extensive reaction networks that represent complex chemical phenomena and processes at an unprecedented depth. We illustrate our methodology for the Bray–Liebhafsky reaction, an oscillating reaction (that is, a chemical clock).
Navigating the intersection of data privacy and machine learning, our approach can offer secure data sharing for the chemical sciences. Utilizing cutting-edge encryption techniques, it ensures privacy while enabling effective machine learning models to operate. This fosters academia-industry collaboration for expediting the discovery of new materials. With implications reaching beyond the chemical sector, into areas like drug development, paving a path for future innovations and sustainable solutions.
We're taking scientific discovery to the next level with DARWIN, a cutting-edge tool that supercharges research in physics, chemistry, and material science. By leveraging vast amounts of data and AI, DARWIN automates the heavy lifting in research, accelerates discoveries, and outperforms traditional models. Our solution seamlessly integrates with your work, transforming how scientific knowledge is applied and unlocking the limitless potential of your projects. Discover the future of natural science with DARWIN!
SDLs consist of multiple stations that perform material synthesis and characterisation tasks. To minimize station downtime, it is practical to run experiments at once in different stages. This, however, introduces delayed feedback, which is known to reduce optimizer performance. Using a simulator, we compare search strategies such as naive expected improvement, 4-mode exploration and asynchronous batching. Our simulation results showcase the trade-off between the asynchronous parallel operation and delayed feedback.
Drug and development of new functionalized materials requires macro-quantities on the gram scale to ensure enough material at the end to enable structure determination and confirmation of the desired target. Serial Nano-Electron Diffraction, with a new method for creating showers of nanocrystals on demand, provides the highest possible structural resolution, using only microgram to nanograms quantities – opening the door to true nanoscale chemistry with accompanying orders of magnitude speed up in the discovery pipeline.
For the development and wider adoption of self-driving labs (SDL), the availability of automated and sustainable manufacturing techniques that allow for accelerated and scalable synthesis of materials is a prerequisite. Conventional wet chemistry techniques are recipe-based and involve multiple steps for the production of nanoparticles and layers. Iterations for conventional synthesis alone usually takes many years, but also the wet chemistry steps hamper integration into an SDL. An alternative approach to these methods is the proposed single-step, dry synthesis and deposition of materials.
Our framework integrates unsupervised learning with chemistry classification, addressing overconfidence in supervised methods. By training class-specific unsupervised models, we calculate likeliness scores to measure sample similarity and avoid overconfident predictions on unseen data. Testing on benchmark datasets reveals comparable performance to supervised methods, with superior detection of out-of-sample samples. Our approach holds potential for reliable predictions in chemistry, particularly in domains where safe classification is vital.
Our research unveils a new area of application for the Hammett equation: the prediction of binding energies of homogeneous catalysts. For such a simple and inexpensive method, the predictions are sufficiently accurate to identify ideal catalysts. We further enhance simplicity and accuracy with combining rules and delta machine learning, additions that can make a significant contribution to accelerating catalyst design.
We are surrounded by structures designed to absorb energy, such as padding in sports equipment or packaging for shipping. Efficient energy absorption can decrease the size and cost of the structures while increasing safety. By using a self-driving lab to perform >12,500 physical experiments, we discovered the most efficient structure for absorbing energy that has ever been observed. This large database of experiments also provides insight into the dichotomy of elastic vs. plastic materials.
We propose a novel generative model for designing molecules with desired properties that assembles retrosynthetically prepared chemical building blocks to improve the synthesizability. The model can efficiently use dozens of thousands of building blocks including unseen blocks for molecule generation, which was not possible in previous models. We demonstrate the real-world applicability of our strategy by generating potential molecules in various material discovery domains.
We present a scheme for the acceleration of chemical space exploration based on chemical reaction networks. Combining fast, but inaccurate quantum mechanical calculations with smart selection strategies and slow, but accurate quantum chemical methods allows for constructing extensive reaction networks that represent complex chemical phenomena and processes at an unprecedented depth. We illustrate our methodology for the Bray–Liebhafsky reaction, an oscillating reaction (that is, a chemical clock).
This study used a self-driving lab to identify structured polymer components that efficiently absorb mechanical energy (K). We hypothesized that pairing components can yield superior K. The complexity of this optimization problem was tackled by computationally screening >12,000 components and implementing an active learning algorithm that fixed one component to test pairings. Besides identifying complex structures that are high efficiency, this research may shed light on how to screen for emergent properties of complex systems.
Here, we use an experiment that isolates high-affinity binding molecules from large diverse combinatorial libraries, called AS-MS, along with machine learning (ML) to guide the discovery of ultra-high affinity binding compounds. From the AS-MS data, the ML provided a “map” that enabled the accurate prediction of high-affinity binding across multiple libraries, natural and non-natural. We hope that this mapping approach will accelerate drug discovery and development.
We report a two-stage data-driven approach to the targeted synthesis of organic photoredox catalysts and the subsequent reaction optimization for metallophotocatalysis, as demonstrated for decarboxylative C(sp3)–C(sp2) cross-coupling of amino acids with aryl halides. The broad relevance of our approach to other areas is enhanced because our approach does not involve any specialized automation or robotics equipment; all of the chemistry is done by hand, and the efficiency stems from the use of Bayesian optimization-led searches.
We introduce Chemcrow, a chemistry agent that integrates 17 expert-designed tools to enhance performance in chemistry-related tasks such as organic synthesis, drug discovery, and materials design. Our evaluation, including both LLM and expert assessments, demonstrates ChemCrow's effectiveness. However, there are potential risks and misuse concerns, which we discuss. ChemCrow bridges the gap between experimental and computational chemistry, benefiting experts and non-experts while fostering scientific advancement.
Traditional ingredients in everyday products such as shampoo are getting eco-conscious consumers in a lather. Machine learning methods can be used to design better molecules & materials; however, they require experimental data to train on. Performing the experiments is typically time, resource and labour-intensive. To alleviate this, a high-throughput liquid formulation workflow using robotics and AI has been developed. This helps accelerate product development and brings us closer to making more eco-friendly products.
Our approach optimizes chemistry experiments with a robot incorporating real-time feedback from a crowd of expert chemists into Bayesian Optimization. The robot uses an interactive website to enable experts to share their hypotheses. This AI-human collaboration allows for more adventurous search space exploration and faster convergence. Our experimental results show that this approach improves BO with expert hypotheses, even in cases where the expert's knowledge is inaccurate and can recover from misleading hypotheses.
Automated recognition of molecular images using AI has recently emerged as an intriguing challenge. Here, we suggest an unsupervised learning model for molecular image recognition by converting the image to a graph representation that enables image reconstruction. By comparing the reconstructed image to the original one, the model will learn the accurate graph representation in an unsupervised manner. The permutation-invariant nature of the model is expected to enhance its accuracy compared to previous ones.
Recent advances in large language models (LLMs) provide powerful tools for predictive modeling on a variety of natural language tasks. Prior work developing Genome-scale Language Models (GenSLMs) demonstrated the potential for LLMs to predict future SARS-CoV-2 variants of concern prior to their emergence by modeling the evolutionary process. We propose to expand this framework by demonstrating how to effectively scale GenSLMs for bacterial genomes, and even further to more complex eukaryotic organisms including yeast and humans.
Analysis bonding patterns in high entropy alloys using EXAFS is a difficult and time consuming task that could greatly benefit from active learning (AL). Like many other potential applications of AL, the complexity of the EXAFS analysis problem complicates direct application of AL. We present our approach combining on-the-fly autoencoders with quantitative analysis based on modern statistical methods.
Integrating evolution algorithms with Bayesian optimization, we enable efficient materials discovery in self driving labs.
An overlooked challenge of operating an autonomous chemistry lab is the high-level orchestration and planning required to efficiently execute multiple workflows (e.g., multiple syntheses) in a way that maximizes throughput. We view this as a scheduling and optimization problem. We have developed a discrete event simulator, “LASSO”, to serve as a digital twin for an automated synthesis platform and an environment within which to test scheduling algorithms.
The Fuel Cell INFormation Ontology (FCINFO) and Fuel Cell Value Chain Ontology (FCVCO) aim to enhance knowledge discovery, data integration, and collaboration in fuel cell research, development, and manufacturing. FCINFO concentrates on fuel cells, their components, materials, and interfaces, while FCVCO promotes knowledge integration, collaboration, and decision-making throughout different fuel cell technology development phases. Both ontologies offer a structured framework for representing concepts and relationships within the fuel cell domain.
We introduce AutoProtocol, a unique framework that uses advanced AI and robotic simulations to automate molecular biology experiments. This system, trained on textual data from sources like Science Exchange, PLOS, and Bio-protocol, generates intricate protocols. It takes into account biologist feedback and error corrections from digital twin simulations, minimizing human error and enhancing reproducibility. AutoProtocol is a significant step towards self-driving molecular biology laboratories, saving time and effort in experiment design and execution.
We're taking scientific discovery to the next level with DARWIN, a cutting-edge tool that supercharges research in physics, chemistry, and material science. By leveraging vast amounts of data and AI, DARWIN automates the heavy lifting in research, accelerates discoveries, and outperforms traditional models. Our solution seamlessly integrates with your work, transforming how scientific knowledge is applied and unlocking the limitless potential of your projects. Discover the future of natural science with DARWIN!
SDLs consist of multiple stations that perform material synthesis and characterisation tasks. To minimize station downtime, it is practical to run experiments at once in different stages. This, however, introduces delayed feedback, which is known to reduce optimizer performance. Using a simulator, we compare search strategies such as naive expected improvement, 4-mode exploration and asynchronous batching. Our simulation results showcase the trade-off between the asynchronous parallel operation and delayed feedback.
We have designed and are prototyping a self-driving laboratory simulation ecosystem in NVIDIA Omniverse and Isaac Sim to enable rapid design of laboratories and of the experiments they perform. Our primary focus is a set of tools that improves access to self-driving laboratory technologies and simplifies interfaces for faster learning how to use the system.
Creating a community-driven ecosystem of open-source hardware designed for laboratory automation solution can help broaden the application space and increase the implementation of self-driving laboratories. Through the fabricable multi-tool motion platform, Jubilee, the project aims to create community-driven resources for the development and preservation of open-hardware ecosystems. These will allow a broader application space and enable a scientific maker movement for more inclusion of automation tools into research environments and educational curricula.
Navigating the intersection of data privacy and machine learning, our approach can offer secure data sharing for the chemical sciences. Utilizing cutting-edge encryption techniques, it ensures privacy while enabling effective machine learning models to operate. This fosters academia-industry collaboration for expediting the discovery of new materials. With implications reaching beyond the chemical sector, into areas like drug development, paving a path for future innovations and sustainable solutions.
This study used a self-driving lab to identify structured polymer components that efficiently absorb mechanical energy (K). We hypothesized that pairing components can yield superior K. The complexity of this optimization problem was tackled by computationally screening >12,000 components and implementing an active learning algorithm that fixed one component to test pairings. Besides identifying complex structures that are high efficiency, this research may shed light on how to screen for emergent properties of complex systems.
We have designed and are prototyping a self-driving laboratory simulation ecosystem in NVIDIA Omniverse and Isaac Sim to enable rapid design of laboratories and of the experiments they perform. Our primary focus is a set of tools that improves access to self-driving laboratory technologies and simplifies interfaces for faster learning how to use the system.
Here, we use an experiment that isolates high-affinity binding molecules from large diverse combinatorial libraries, called AS-MS, along with machine learning (ML) to guide the discovery of ultra-high affinity binding compounds. From the AS-MS data, the ML provided a “map” that enabled the accurate prediction of high-affinity binding across multiple libraries, natural and non-natural. We hope that this mapping approach will accelerate drug discovery and development.
We report a two-stage data-driven approach to the targeted synthesis of organic photoredox catalysts and the subsequent reaction optimization for metallophotocatalysis, as demonstrated for decarboxylative C(sp3)–C(sp2) cross-coupling of amino acids with aryl halides. The broad relevance of our approach to other areas is enhanced because our approach does not involve any specialized automation or robotics equipment; all of the chemistry is done by hand, and the efficiency stems from the use of Bayesian optimization-led searches.
We introduce Chemcrow, a chemistry agent that integrates 17 expert-designed tools to enhance performance in chemistry-related tasks such as organic synthesis, drug discovery, and materials design. Our evaluation, including both LLM and expert assessments, demonstrates ChemCrow's effectiveness. However, there are potential risks and misuse concerns, which we discuss. ChemCrow bridges the gap between experimental and computational chemistry, benefiting experts and non-experts while fostering scientific advancement.
Traditional ingredients in everyday products such as shampoo are getting eco-conscious consumers in a lather. Machine learning methods can be used to design better molecules & materials; however, they require experimental data to train on. Performing the experiments is typically time, resource and labour-intensive. To alleviate this, a high-throughput liquid formulation workflow using robotics and AI has been developed. This helps accelerate product development and brings us closer to making more eco-friendly products.
Drug and development of new functionalized materials requires macro-quantities on the gram scale to ensure enough material at the end to enable structure determination and confirmation of the desired target. Serial Nano-Electron Diffraction, with a new method for creating showers of nanocrystals on demand, provides the highest possible structural resolution, using only microgram to nanograms quantities – opening the door to true nanoscale chemistry with accompanying orders of magnitude speed up in the discovery pipeline.
Our approach optimizes chemistry experiments with a robot incorporating real-time feedback from a crowd of expert chemists into Bayesian Optimization. The robot uses an interactive website to enable experts to share their hypotheses. This AI-human collaboration allows for more adventurous search space exploration and faster convergence. Our experimental results show that this approach improves BO with expert hypotheses, even in cases where the expert's knowledge is inaccurate and can recover from misleading hypotheses.
For the development and wider adoption of self-driving labs (SDL), the availability of automated and sustainable manufacturing techniques that allow for accelerated and scalable synthesis of materials is a prerequisite. Conventional wet chemistry techniques are recipe-based and involve multiple steps for the production of nanoparticles and layers. Iterations for conventional synthesis alone usually takes many years, but also the wet chemistry steps hamper integration into an SDL. An alternative approach to these methods is the proposed single-step, dry synthesis and deposition of materials.
Automated recognition of molecular images using AI has recently emerged as an intriguing challenge. Here, we suggest an unsupervised learning model for molecular image recognition by converting the image to a graph representation that enables image reconstruction. By comparing the reconstructed image to the original one, the model will learn the accurate graph representation in an unsupervised manner. The permutation-invariant nature of the model is expected to enhance its accuracy compared to previous ones.
Recent advances in large language models (LLMs) provide powerful tools for predictive modeling on a variety of natural language tasks. Prior work developing Genome-scale Language Models (GenSLMs) demonstrated the potential for LLMs to predict future SARS-CoV-2 variants of concern prior to their emergence by modeling the evolutionary process. We propose to expand this framework by demonstrating how to effectively scale GenSLMs for bacterial genomes, and even further to more complex eukaryotic organisms including yeast and humans.
Analysis bonding patterns in high entropy alloys using EXAFS is a difficult and time consuming task that could greatly benefit from active learning (AL). Like many other potential applications of AL, the complexity of the EXAFS analysis problem complicates direct application of AL. We present our approach combining on-the-fly autoencoders with quantitative analysis based on modern statistical methods.
Our framework integrates unsupervised learning with chemistry classification, addressing overconfidence in supervised methods. By training class-specific unsupervised models, we calculate likeliness scores to measure sample similarity and avoid overconfident predictions on unseen data. Testing on benchmark datasets reveals comparable performance to supervised methods, with superior detection of out-of-sample samples. Our approach holds potential for reliable predictions in chemistry, particularly in domains where safe classification is vital.
Integrating evolution algorithms with Bayesian optimization, we enable efficient materials discovery in self driving labs.
An overlooked challenge of operating an autonomous chemistry lab is the high-level orchestration and planning required to efficiently execute multiple workflows (e.g., multiple syntheses) in a way that maximizes throughput. We view this as a scheduling and optimization problem. We have developed a discrete event simulator, “LASSO”, to serve as a digital twin for an automated synthesis platform and an environment within which to test scheduling algorithms.
Creating a community-driven ecosystem of open-source hardware designed for laboratory automation solution can help broaden the application space and increase the implementation of self-driving laboratories. Through the fabricable multi-tool motion platform, Jubilee, the project aims to create community-driven resources for the development and preservation of open-hardware ecosystems. These will allow a broader application space and enable a scientific maker movement for more inclusion of automation tools into research environments and educational curricula.
Our research unveils a new area of application for the Hammett equation: the prediction of binding energies of homogeneous catalysts. For such a simple and inexpensive method, the predictions are sufficiently accurate to identify ideal catalysts. We further enhance simplicity and accuracy with combining rules and delta machine learning, additions that can make a significant contribution to accelerating catalyst design.
We are surrounded by structures designed to absorb energy, such as padding in sports equipment or packaging for shipping. Efficient energy absorption can decrease the size and cost of the structures while increasing safety. By using a self-driving lab to perform >12,500 physical experiments, we discovered the most efficient structure for absorbing energy that has ever been observed. This large database of experiments also provides insight into the dichotomy of elastic vs. plastic materials.
The Fuel Cell INFormation Ontology (FCINFO) and Fuel Cell Value Chain Ontology (FCVCO) aim to enhance knowledge discovery, data integration, and collaboration in fuel cell research, development, and manufacturing. FCINFO concentrates on fuel cells, their components, materials, and interfaces, while FCVCO promotes knowledge integration, collaboration, and decision-making throughout different fuel cell technology development phases. Both ontologies offer a structured framework for representing concepts and relationships within the fuel cell domain.
We introduce AutoProtocol, a unique framework that uses advanced AI and robotic simulations to automate molecular biology experiments. This system, trained on textual data from sources like Science Exchange, PLOS, and Bio-protocol, generates intricate protocols. It takes into account biologist feedback and error corrections from digital twin simulations, minimizing human error and enhancing reproducibility. AutoProtocol is a significant step towards self-driving molecular biology laboratories, saving time and effort in experiment design and execution.
We propose a novel generative model for designing molecules with desired properties that assembles retrosynthetically prepared chemical building blocks to improve the synthesizability. The model can efficiently use dozens of thousands of building blocks including unseen blocks for molecule generation, which was not possible in previous models. We demonstrate the real-world applicability of our strategy by generating potential molecules in various material discovery domains.
We present a scheme for the acceleration of chemical space exploration based on chemical reaction networks. Combining fast, but inaccurate quantum mechanical calculations with smart selection strategies and slow, but accurate quantum chemical methods allows for constructing extensive reaction networks that represent complex chemical phenomena and processes at an unprecedented depth. We illustrate our methodology for the Bray–Liebhafsky reaction, an oscillating reaction (that is, a chemical clock).
Navigating the intersection of data privacy and machine learning, our approach can offer secure data sharing for the chemical sciences. Utilizing cutting-edge encryption techniques, it ensures privacy while enabling effective machine learning models to operate. This fosters academia-industry collaboration for expediting the discovery of new materials. With implications reaching beyond the chemical sector, into areas like drug development, paving a path for future innovations and sustainable solutions.
We're taking scientific discovery to the next level with DARWIN, a cutting-edge tool that supercharges research in physics, chemistry, and material science. By leveraging vast amounts of data and AI, DARWIN automates the heavy lifting in research, accelerates discoveries, and outperforms traditional models. Our solution seamlessly integrates with your work, transforming how scientific knowledge is applied and unlocking the limitless potential of your projects. Discover the future of natural science with DARWIN!
SDLs consist of multiple stations that perform material synthesis and characterisation tasks. To minimize station downtime, it is practical to run experiments at once in different stages. This, however, introduces delayed feedback, which is known to reduce optimizer performance. Using a simulator, we compare search strategies such as naive expected improvement, 4-mode exploration and asynchronous batching. Our simulation results showcase the trade-off between the asynchronous parallel operation and delayed feedback.