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.
Most procedure data in organic chemistry is represented by unstructured text, which limits its use in data-driven approaches that rely on structured data. We fined-tuned open source large language models to extract structured JSON from free text synthesis procedures, which correctly captured 93% of reactant identities as well as their associated properties.
The Pan-Canadian Chemical Library project aims to enable academic chemists in Canada to create custom combinatorial libraries for virtual screening. As part of a pilot study to evaluate the feasibility of the project, we worked in close collaboration with chemists across Canada to identify innovative chemical reactions, allowing the generation of over 115 billion compounds, including 179 million that we estimate can be synthesized quickly at low cost. More information on https://pccl.thesgc.org/
Eleven ML algorithms were trained to predict fractional drug release for LAIs. Model interpretation steps were utilized to excerpt learned knowledge on these drug-polymer systems. This extracted knowledge was used to identify design criteria for fast-release and slow-release poly(lactide-co-glycolide) (PLGA) formulations. These LAIs were then prepared using an oil-in-water (o/w) emulsion method and their in vitro release was characterized.
To accelerate the development of optoelectronic thin films with desired characteristics, we developed a high throughput experimentation platform combined with hyperspectral imaging system and automated probe station for measuring spectral and electrical properties respectively. We implemented automated extraction and visualization of spectral/electrical features to optimize thin films effectively.
Materials graph neural networks are employed to construct machine learning force fields for molecular dynamics simulations. These networks capture interatomic interactions and learn the underlying physics by representing molecules or materials as graphs. This approach reduces computational costs, enables the study of larger systems, and accurately reproduces structural energies and trajectories. It has the potential to accelerate materials discovery and facilitate the design of new materials with desired properties.
This research developed a machine learning model to predict the voltage of battery cathode materials, enabling the rapid screening of new materials. The model was trained using the Materials Project dataset, and its predictive power was improved by adding additional data points, including materials with unfavorable lithium binding. The model was used to identify 12 viable new cathode materials.
How can you learn to predict peptide properties without negative examples? This often happens when trying to analyze outputs from screening results. Peptide screening usually gives positive examples, which makes it difficult to train a classifier. In this work, we only use the known positive examples for training and make predictions on properties of new peptide sequences.
By combining innovative spark nano-mixing and rapid sintering techniques, we accelerated solid state synthesis hindered by slow solid diffusion and inadequate mixing. We demonstrate alloying of Cu/Ni from as low as 100 °C and fabrication of high-purity GeTe less than 1s. Embrace a new frontier of possibilities for high-throughput production of multi-component solid materials with our transformative approach.
The accumulation of polar pesticides in the environment such as glyphosate (Round Up) is a long-term threat for clean and accessible drinking water and for aquatic life. Synthetic reagents are necessary to facilitate highly sensitive analytical methods for monitoring contaminants in the environment. We will develop a machine learning algorithm for the inverse design of synthetic reagents for the sustainable and accurate quantification of glyphosate in water samples.
The hydrogen economy will rely on the safe and efficient storage of hydrogen. Machine learning and simulations are employed to discover new metal alloys that can absorb and desorb hydrogen at room temperature. Optimal alloys are discovered, and insights are obtained to reveal the underlying material properties that determine the hydrogen storage performance.
The properties of organic molecular crystals are highly dependent not only on the molecules that form the crystal structure but also on their arrangement. Therefore, a method has been developed incorporating crystal structure prediction into a directed search through chemical space, in the form of an evolutionary algorithm. This enables the sampled molecules to be evaluated based upon their more representative likely material properties and has been demonstrated on the organic semiconductor search space.
Closed-loop approaches, where machine learning generated hypotheses guide experiments, hold promise for rapid materials discovery. Here we demonstrate a computational workflow that autonomously searches for catalyst materials towards decarbonizing the agricultural industry through electrification. In contrast to previous efforts, our workflow seeks to simultaneously optimize multiple key target metrics on-the-fly rather than focusing on any single property relevant to catalyst performance. We further show how unearthed candidate systems can be ranked for prioritizing followup investigations.
Preparing a desired chemical compound requires identifying possible synthetic routes and methods. Reaction scopes can inform chemists about the reactions' limitations and capabilities; yet, applying the same reaction to a reactant that was never tested before is not trivial. Employing a trial-and-error method based on intuition can be expensive and frustrating; therefore, in this work we aimed to harness the power of classification to facilitate the choice of one set of reaction conditions over another.
We introduce an automated data-extraction pipeline from peer-reviewed literature, which may accelerate scientific discoveries by generating reliable large-scale datasets for statistical learning. This is exemplified by curating a dataset of experimentally-measured oxidation potentials of organic molecules to train multiple machine learning models, which are used to predict oxidation potentials for 132k molecules. We envision the usefulness of our approach extending beyond chemistry by uncovering hidden treasures of data in the literature of various fields.
By combining robotics and machine learning, we've developed an automated experiment named "Clio" and an intelligent experiment planner called "Dragonfly." In just two days and forty-two experiments, our system identified six fast-charging non-aqueous electrolyte solutions, six times faster than random searches. These solutions were validated in real-world tests and showed improved fast-charging capabilities in a pouch cell configuration. Accelerate battery optimization and accelerate electrification with our technology!
Discover the future of complex fluid experiments with our groundbreaking technology. Our method enables automated sample exchange using a pendant drop, eliminating manual processes and contamination risks. With high precision and repeatability, our robotic system prepares liquid samples with tailored compositions. The entire life cycle, from preparation to disposal, is fully automated. Experience seamless integration with our Python-based solution, empowering AI-driven experiments for autonomous material design. Revolutionize your research with our collaborative development platform.
Discovering new solid-state materials via Materials Acceleration Platforms (MAPs) is challenging due to the lack of a general method to rapidly synthesize bulk materials. Here, we present rapid solid-state synthesis techniques via laser heating and rapid joule heating of ternary chalcogenide semiconductors, achieving phase-pure crystalline materials synthesized in the milligram scale in as little as 15 seconds, accelerating the solid-state reaction process by a factor of >100 relative to the traditional shake-and-bake route.
Progress in data-driven approaches and machine-guided optimizations has accelerated the discovery of new molecules with desired functions and properties. High-fidelity experimental data-driven machine learning (ML) predictions were validated with high accuracy in three case studies.
In our study, we investigate various MOF representations using Bayesian Optimization. With a training set of under 120 and judicious feature selection, we identified the top-performing MOFs from a pool exceeding 15,000. Our research will present a structured blueprint for feature selection, aiming for efficient optimization of desired properties by self-driving labs.
Application-specific materials-by-design is a long-standing challenge due to the need for capturing the process-microstructure-property relations. The efficient identification of microstructures inversely from target mechanical properties is intractable because of microstructures' complexity. Here, we provide an end-to-end framework that tackles both forward and inverse predictions to streamline materials-by-design. Using advanced deep-learning and genetic algorithm, our framework exhibits promising potential in cutting down the time needed from target mechanical properties directly to desired material microstructure.
Rapid, automated access to small molecules could enable more facile access to materials, thus leading to democratized discovery of molecular function.
Transparent object perception is a crucial skill for applications such as robot manipulation in household and laboratory settings. We propose MVTrans, a multiview multitask perception method, and create Syn-TODD, a large-scale transparent object detection dataset. Which surpasses existing RGB-D and stereo-based methods for handling transparent object perception tasks.
Our work generalizes structure-based ligand design by employing protein-ligand interactions in a 3D molecular generative model. We propose an interaction-focused strategy that captures the surrounding pocket environment, precisely navigating the ligand generation by pursuing the generalizable geometric pattern of intermolecular interactions. Designed ligands achieve stable binding by forming favorable interactions regardless of the target protein. Finally, we emphasize broad applicability through various ligand-designing tasks where specific interactions play a significant role.
Self-emulsifying drug delivery systems (SEDDS) are an advanced formulation strategy used to improve the oral absorption of poorly water-soluble drugs. Designing SEDDS formulations is time and resource intensive based on current approaches, particularly for selecting the composition of excipients. Given their longstanding history, there is an opportunity to leverage information from previously reported SEDDS. We created a comprehensive SEDDS dataset and modelled formulation properties, which may ultimately assist the formulation development process.
The Pan-Canadian Chemical Library project aims to enable academic chemists in Canada to create custom combinatorial libraries for virtual screening. As part of a pilot study to evaluate the feasibility of the project, we worked in close collaboration with chemists across Canada to identify innovative chemical reactions, allowing the generation of over 115 billion compounds, including 179 million that we estimate can be synthesized quickly at low cost. More information on https://pccl.thesgc.org/
Materials graph neural networks are employed to construct machine learning force fields for molecular dynamics simulations. These networks capture interatomic interactions and learn the underlying physics by representing molecules or materials as graphs. This approach reduces computational costs, enables the study of larger systems, and accurately reproduces structural energies and trajectories. It has the potential to accelerate materials discovery and facilitate the design of new materials with desired properties.
This research developed a machine learning model to predict the voltage of battery cathode materials, enabling the rapid screening of new materials. The model was trained using the Materials Project dataset, and its predictive power was improved by adding additional data points, including materials with unfavorable lithium binding. The model was used to identify 12 viable new cathode materials.
How can you learn to predict peptide properties without negative examples? This often happens when trying to analyze outputs from screening results. Peptide screening usually gives positive examples, which makes it difficult to train a classifier. In this work, we only use the known positive examples for training and make predictions on properties of new peptide sequences.
By combining innovative spark nano-mixing and rapid sintering techniques, we accelerated solid state synthesis hindered by slow solid diffusion and inadequate mixing. We demonstrate alloying of Cu/Ni from as low as 100 °C and fabrication of high-purity GeTe less than 1s. Embrace a new frontier of possibilities for high-throughput production of multi-component solid materials with our transformative approach.
The accumulation of polar pesticides in the environment such as glyphosate (Round Up) is a long-term threat for clean and accessible drinking water and for aquatic life. Synthetic reagents are necessary to facilitate highly sensitive analytical methods for monitoring contaminants in the environment. We will develop a machine learning algorithm for the inverse design of synthetic reagents for the sustainable and accurate quantification of glyphosate in water samples.
The hydrogen economy will rely on the safe and efficient storage of hydrogen. Machine learning and simulations are employed to discover new metal alloys that can absorb and desorb hydrogen at room temperature. Optimal alloys are discovered, and insights are obtained to reveal the underlying material properties that determine the hydrogen storage performance.
The properties of organic molecular crystals are highly dependent not only on the molecules that form the crystal structure but also on their arrangement. Therefore, a method has been developed incorporating crystal structure prediction into a directed search through chemical space, in the form of an evolutionary algorithm. This enables the sampled molecules to be evaluated based upon their more representative likely material properties and has been demonstrated on the organic semiconductor search space.
Closed-loop approaches, where machine learning generated hypotheses guide experiments, hold promise for rapid materials discovery. Here we demonstrate a computational workflow that autonomously searches for catalyst materials towards decarbonizing the agricultural industry through electrification. In contrast to previous efforts, our workflow seeks to simultaneously optimize multiple key target metrics on-the-fly rather than focusing on any single property relevant to catalyst performance. We further show how unearthed candidate systems can be ranked for prioritizing followup investigations.
Preparing a desired chemical compound requires identifying possible synthetic routes and methods. Reaction scopes can inform chemists about the reactions' limitations and capabilities; yet, applying the same reaction to a reactant that was never tested before is not trivial. Employing a trial-and-error method based on intuition can be expensive and frustrating; therefore, in this work we aimed to harness the power of classification to facilitate the choice of one set of reaction conditions over another.
Discovering new solid-state materials via Materials Acceleration Platforms (MAPs) is challenging due to the lack of a general method to rapidly synthesize bulk materials. Here, we present rapid solid-state synthesis techniques via laser heating and rapid joule heating of ternary chalcogenide semiconductors, achieving phase-pure crystalline materials synthesized in the milligram scale in as little as 15 seconds, accelerating the solid-state reaction process by a factor of >100 relative to the traditional shake-and-bake route.
Progress in data-driven approaches and machine-guided optimizations has accelerated the discovery of new molecules with desired functions and properties. High-fidelity experimental data-driven machine learning (ML) predictions were validated with high accuracy in three case studies.
In our study, we investigate various MOF representations using Bayesian Optimization. With a training set of under 120 and judicious feature selection, we identified the top-performing MOFs from a pool exceeding 15,000. Our research will present a structured blueprint for feature selection, aiming for efficient optimization of desired properties by self-driving labs.
Application-specific materials-by-design is a long-standing challenge due to the need for capturing the process-microstructure-property relations. The efficient identification of microstructures inversely from target mechanical properties is intractable because of microstructures' complexity. Here, we provide an end-to-end framework that tackles both forward and inverse predictions to streamline materials-by-design. Using advanced deep-learning and genetic algorithm, our framework exhibits promising potential in cutting down the time needed from target mechanical properties directly to desired material microstructure.
Rapid, automated access to small molecules could enable more facile access to materials, thus leading to democratized discovery of molecular function.
Our work generalizes structure-based ligand design by employing protein-ligand interactions in a 3D molecular generative model. We propose an interaction-focused strategy that captures the surrounding pocket environment, precisely navigating the ligand generation by pursuing the generalizable geometric pattern of intermolecular interactions. Designed ligands achieve stable binding by forming favorable interactions regardless of the target protein. Finally, we emphasize broad applicability through various ligand-designing tasks where specific interactions play a significant role.
Self-emulsifying drug delivery systems (SEDDS) are an advanced formulation strategy used to improve the oral absorption of poorly water-soluble drugs. Designing SEDDS formulations is time and resource intensive based on current approaches, particularly for selecting the composition of excipients. Given their longstanding history, there is an opportunity to leverage information from previously reported SEDDS. We created a comprehensive SEDDS dataset and modelled formulation properties, which may ultimately assist the formulation development process.
Most procedure data in organic chemistry is represented by unstructured text, which limits its use in data-driven approaches that rely on structured data. We fined-tuned open source large language models to extract structured JSON from free text synthesis procedures, which correctly captured 93% of reactant identities as well as their associated properties.
To accelerate the development of optoelectronic thin films with desired characteristics, we developed a high throughput experimentation platform combined with hyperspectral imaging system and automated probe station for measuring spectral and electrical properties respectively. We implemented automated extraction and visualization of spectral/electrical features to optimize thin films effectively.
We introduce an automated data-extraction pipeline from peer-reviewed literature, which may accelerate scientific discoveries by generating reliable large-scale datasets for statistical learning. This is exemplified by curating a dataset of experimentally-measured oxidation potentials of organic molecules to train multiple machine learning models, which are used to predict oxidation potentials for 132k molecules. We envision the usefulness of our approach extending beyond chemistry by uncovering hidden treasures of data in the literature of various fields.
By combining robotics and machine learning, we've developed an automated experiment named "Clio" and an intelligent experiment planner called "Dragonfly." In just two days and forty-two experiments, our system identified six fast-charging non-aqueous electrolyte solutions, six times faster than random searches. These solutions were validated in real-world tests and showed improved fast-charging capabilities in a pouch cell configuration. Accelerate battery optimization and accelerate electrification with our technology!
Discover the future of complex fluid experiments with our groundbreaking technology. Our method enables automated sample exchange using a pendant drop, eliminating manual processes and contamination risks. With high precision and repeatability, our robotic system prepares liquid samples with tailored compositions. The entire life cycle, from preparation to disposal, is fully automated. Experience seamless integration with our Python-based solution, empowering AI-driven experiments for autonomous material design. Revolutionize your research with our collaborative development platform.
Transparent object perception is a crucial skill for applications such as robot manipulation in household and laboratory settings. We propose MVTrans, a multiview multitask perception method, and create Syn-TODD, a large-scale transparent object detection dataset. Which surpasses existing RGB-D and stereo-based methods for handling transparent object perception tasks.
Most procedure data in organic chemistry is represented by unstructured text, which limits its use in data-driven approaches that rely on structured data. We fined-tuned open source large language models to extract structured JSON from free text synthesis procedures, which correctly captured 93% of reactant identities as well as their associated properties.
The Pan-Canadian Chemical Library project aims to enable academic chemists in Canada to create custom combinatorial libraries for virtual screening. As part of a pilot study to evaluate the feasibility of the project, we worked in close collaboration with chemists across Canada to identify innovative chemical reactions, allowing the generation of over 115 billion compounds, including 179 million that we estimate can be synthesized quickly at low cost. More information on https://pccl.thesgc.org/
Eleven ML algorithms were trained to predict fractional drug release for LAIs. Model interpretation steps were utilized to excerpt learned knowledge on these drug-polymer systems. This extracted knowledge was used to identify design criteria for fast-release and slow-release poly(lactide-co-glycolide) (PLGA) formulations. These LAIs were then prepared using an oil-in-water (o/w) emulsion method and their in vitro release was characterized.
To accelerate the development of optoelectronic thin films with desired characteristics, we developed a high throughput experimentation platform combined with hyperspectral imaging system and automated probe station for measuring spectral and electrical properties respectively. We implemented automated extraction and visualization of spectral/electrical features to optimize thin films effectively.
Materials graph neural networks are employed to construct machine learning force fields for molecular dynamics simulations. These networks capture interatomic interactions and learn the underlying physics by representing molecules or materials as graphs. This approach reduces computational costs, enables the study of larger systems, and accurately reproduces structural energies and trajectories. It has the potential to accelerate materials discovery and facilitate the design of new materials with desired properties.
This research developed a machine learning model to predict the voltage of battery cathode materials, enabling the rapid screening of new materials. The model was trained using the Materials Project dataset, and its predictive power was improved by adding additional data points, including materials with unfavorable lithium binding. The model was used to identify 12 viable new cathode materials.
How can you learn to predict peptide properties without negative examples? This often happens when trying to analyze outputs from screening results. Peptide screening usually gives positive examples, which makes it difficult to train a classifier. In this work, we only use the known positive examples for training and make predictions on properties of new peptide sequences.
By combining innovative spark nano-mixing and rapid sintering techniques, we accelerated solid state synthesis hindered by slow solid diffusion and inadequate mixing. We demonstrate alloying of Cu/Ni from as low as 100 °C and fabrication of high-purity GeTe less than 1s. Embrace a new frontier of possibilities for high-throughput production of multi-component solid materials with our transformative approach.
The accumulation of polar pesticides in the environment such as glyphosate (Round Up) is a long-term threat for clean and accessible drinking water and for aquatic life. Synthetic reagents are necessary to facilitate highly sensitive analytical methods for monitoring contaminants in the environment. We will develop a machine learning algorithm for the inverse design of synthetic reagents for the sustainable and accurate quantification of glyphosate in water samples.
The hydrogen economy will rely on the safe and efficient storage of hydrogen. Machine learning and simulations are employed to discover new metal alloys that can absorb and desorb hydrogen at room temperature. Optimal alloys are discovered, and insights are obtained to reveal the underlying material properties that determine the hydrogen storage performance.
The properties of organic molecular crystals are highly dependent not only on the molecules that form the crystal structure but also on their arrangement. Therefore, a method has been developed incorporating crystal structure prediction into a directed search through chemical space, in the form of an evolutionary algorithm. This enables the sampled molecules to be evaluated based upon their more representative likely material properties and has been demonstrated on the organic semiconductor search space.
Closed-loop approaches, where machine learning generated hypotheses guide experiments, hold promise for rapid materials discovery. Here we demonstrate a computational workflow that autonomously searches for catalyst materials towards decarbonizing the agricultural industry through electrification. In contrast to previous efforts, our workflow seeks to simultaneously optimize multiple key target metrics on-the-fly rather than focusing on any single property relevant to catalyst performance. We further show how unearthed candidate systems can be ranked for prioritizing followup investigations.
Preparing a desired chemical compound requires identifying possible synthetic routes and methods. Reaction scopes can inform chemists about the reactions' limitations and capabilities; yet, applying the same reaction to a reactant that was never tested before is not trivial. Employing a trial-and-error method based on intuition can be expensive and frustrating; therefore, in this work we aimed to harness the power of classification to facilitate the choice of one set of reaction conditions over another.
We introduce an automated data-extraction pipeline from peer-reviewed literature, which may accelerate scientific discoveries by generating reliable large-scale datasets for statistical learning. This is exemplified by curating a dataset of experimentally-measured oxidation potentials of organic molecules to train multiple machine learning models, which are used to predict oxidation potentials for 132k molecules. We envision the usefulness of our approach extending beyond chemistry by uncovering hidden treasures of data in the literature of various fields.
By combining robotics and machine learning, we've developed an automated experiment named "Clio" and an intelligent experiment planner called "Dragonfly." In just two days and forty-two experiments, our system identified six fast-charging non-aqueous electrolyte solutions, six times faster than random searches. These solutions were validated in real-world tests and showed improved fast-charging capabilities in a pouch cell configuration. Accelerate battery optimization and accelerate electrification with our technology!
Discover the future of complex fluid experiments with our groundbreaking technology. Our method enables automated sample exchange using a pendant drop, eliminating manual processes and contamination risks. With high precision and repeatability, our robotic system prepares liquid samples with tailored compositions. The entire life cycle, from preparation to disposal, is fully automated. Experience seamless integration with our Python-based solution, empowering AI-driven experiments for autonomous material design. Revolutionize your research with our collaborative development platform.
Discovering new solid-state materials via Materials Acceleration Platforms (MAPs) is challenging due to the lack of a general method to rapidly synthesize bulk materials. Here, we present rapid solid-state synthesis techniques via laser heating and rapid joule heating of ternary chalcogenide semiconductors, achieving phase-pure crystalline materials synthesized in the milligram scale in as little as 15 seconds, accelerating the solid-state reaction process by a factor of >100 relative to the traditional shake-and-bake route.
Progress in data-driven approaches and machine-guided optimizations has accelerated the discovery of new molecules with desired functions and properties. High-fidelity experimental data-driven machine learning (ML) predictions were validated with high accuracy in three case studies.
In our study, we investigate various MOF representations using Bayesian Optimization. With a training set of under 120 and judicious feature selection, we identified the top-performing MOFs from a pool exceeding 15,000. Our research will present a structured blueprint for feature selection, aiming for efficient optimization of desired properties by self-driving labs.
Application-specific materials-by-design is a long-standing challenge due to the need for capturing the process-microstructure-property relations. The efficient identification of microstructures inversely from target mechanical properties is intractable because of microstructures' complexity. Here, we provide an end-to-end framework that tackles both forward and inverse predictions to streamline materials-by-design. Using advanced deep-learning and genetic algorithm, our framework exhibits promising potential in cutting down the time needed from target mechanical properties directly to desired material microstructure.
Rapid, automated access to small molecules could enable more facile access to materials, thus leading to democratized discovery of molecular function.
Transparent object perception is a crucial skill for applications such as robot manipulation in household and laboratory settings. We propose MVTrans, a multiview multitask perception method, and create Syn-TODD, a large-scale transparent object detection dataset. Which surpasses existing RGB-D and stereo-based methods for handling transparent object perception tasks.
Our work generalizes structure-based ligand design by employing protein-ligand interactions in a 3D molecular generative model. We propose an interaction-focused strategy that captures the surrounding pocket environment, precisely navigating the ligand generation by pursuing the generalizable geometric pattern of intermolecular interactions. Designed ligands achieve stable binding by forming favorable interactions regardless of the target protein. Finally, we emphasize broad applicability through various ligand-designing tasks where specific interactions play a significant role.
Self-emulsifying drug delivery systems (SEDDS) are an advanced formulation strategy used to improve the oral absorption of poorly water-soluble drugs. Designing SEDDS formulations is time and resource intensive based on current approaches, particularly for selecting the composition of excipients. Given their longstanding history, there is an opportunity to leverage information from previously reported SEDDS. We created a comprehensive SEDDS dataset and modelled formulation properties, which may ultimately assist the formulation development process.