Posters will be displayed at each of our venues throughout the conference.
Bahen Centre for Information Technology
Myhal Centre for Engineering Innovation and Entrepreneurship
Posters will be displayed at each of our venues throughout the conference.
CrystaLLM is a large language model of the Crystallographic Information File (CIF) format that is capable of reliably generating correct CIF syntax, physically sensible parameters, and plausible crystal structures for many classes of inorganic compounds. It promises to be a useful tool for tasks requiring generative modelling of inorganic crystal structures.
Determining the structure of organic molecules is necessary for identifying unknown compounds and confirming synthesis products. Rotational spectroscopy provides rich but incomplete information about the 3D structure of molecules. We present two methods, a genetic algorithm, and an equivariant diffusion model, to predict the complete molecular structure starting from this incomplete information. Strong performance of the diffusion model points towards the feasibility of identifying molecules from just rotational spectroscopy and molecular formula.
Polyhydroxyalkanoates (PHAs) have drawn attention as sustainable replacements for petroleum-based plastics as the need for alternatives to these materials has increased. We represent Language models for PHA design (LaMP), fine-tuning state-of-the-art large language models to learn PHAs’ molecular structure-property relationship , and further effectively discover new PHAs based on its targeted properties for specific applications.
Self-driving labs (SDLs) powered by machine learning and automation are emerging technologies that can help overcome the time-consuming nature of electrochemical testing. By developing open-source tools for SDLs, we aim to make them accessible to researchers and educators, facilitating the rapid exploration and optimization of electroactive materials for various applications.
We created a data management system to optimize the use of R&D data in hydrogen technology. The underlying data model is highly flexible in representing fabrication workflows, and uses ontologies to create highly connected knowledge graphs. The data structure allows the implementation of large-language models to increase the findability of data.
To enhance efficiency and reduce costs in chemistry labs, we propose a framework that integrates versatile robotic manipulators with various instruments commonly found in chemistry labs. Our approach successfully plans chemistry experiments using natural language, perceives transparent objects, simulates multi-stage manipulations, and executes robot tasks with safety considerations.
The discovery and optimization of efficient and stable catalysts for the acidic oxygen evolution reaction (OER) are pivotal for advancing clean and sustainable energy sources. In our study, we offer a comprehensive analysis of our findings on the design and implementation of a self-driving laboratory for electrodeposited OER catalysts. We address the challenges posed by experimental variability and present a robust optimization framework. Our work provides valuable insights and considerations for future electrochemical self-driving laboratories.
Graft injury affects over 30% of liver transplant recipients and the exact cause of injury can only be identified currently with a liver biopsy. Transplant recipients can develop graft injury due to various etiologies, such as T-cell mediated rejection (TCMR) and non-alcoholic steatohepatitis-LT (NASH-LT). We developed a machine learning tool, CleVER-LG, leveraging clinical variables and methylation patterns on circulating DNA in plasma as a non-invasive diagnostic tool of graft pathology.
Fused Deposition Modeling (FDM) is a popular and cost-effective 3D printing technology. However, calibrating FDM printers to new feedstocks can be a time-consuming trial and error process that acts as a barrier to new users. To address this, we developed a system for autonomous calibration of entry-level FDM printers using computer vision and metaheuristic methods. Results demonstrate that the system can autonomously calibrate printers to achieve submillimeter dimensional accuracy, surpassing the printer's published tolerance.
Optimizing 3D printing parameters for new materials is a very time consuming task, so we have developed an automated workflow involving pellet-based 3D printers, a collaborative robot, camera vision, and custom-designed lab equipment based on Arduino with the objective of automating all the parts of the experiment. In each iteration, we print and evaluate the quality of the material, and our Artificial Intelligence algorithm learns from the results and plans the next experiments.
The task of predicting whether a material can be made in a lab is difficult and depends on many factors. Previous works do not use thermodynamic stability, a relevant factor, when approaching this task. Thus, we tried using stability among other material descriptors to teach a machine learning model to predict synthesizability. Our model is accurate and identifies over 100 new materials as synthesizable.
Synthesis is the major bottleneck in the design of new inorganic materials. We developed a data-driven algorithm to rapidly design synthesis recipes that are faster, more energy efficient, and more selective than conventional approaches. When coupled with automated laboratory efforts, our approach promises to accelerate the time from lab to market for novel energy materials.
PIGNet2: a "versatile" deep learning-based protein-ligand interaction prediction model comprising the potential in both lead optimization and virtual screening
Polymers are at the core of many advances in various fields: transportation, food packaging, energy storage, constructions. Current methods for selecting the proper materials rely on knowledge of polymer properties. Our objective with the Pyridine project is to deploy an artificial intelligence-based electronic interface that allows its user to input a set of targeted properties and the interface would output the optimal polymers to use, greatly accelerating R&D involving polymers.
Making materials is often the bottleneck in (molecular) materials discovery – therefore, parallelizing synthesis can allow for significant acceleration. However, this necessitates an experiment planning AI that accounts for variable throughput or delayed experimental responses. Here, we describe an adaptive framework for asynchronous, parallel optimization, integrating multiple experiment “threads” at different locations. We demonstrate the ability to orchestrate a delocalized campaign towards organic solid-state lasing materials – showcasing the next step towards democratized materials discovery.
In this work we present the Workflow Execution Interface (WEI). Our presentation will explore how to optimize workflows, improve instrument integration, and facilitate the use of high-performance computing, thus laying the groundwork for autonomous discovery. We will present our progress integrating several different laboratories at Argonne National Laboratory and introduce the Rapid Prototyping Laboratory (RPL) as a student-friendly training ground for self-driving laboratories.
Our work developed an open-source tool that assists Electrochemical Impedance Spectroscopy (EIS) analysis, which is a crucial technique widely used in analyzing electrochemical materials. By automatically proposing statistically plausible circuit models, AutoEIS streamlines data interpretation without requiring in-depth knowledge of electrochemical systems. This generalized automated approach enhances the efficiency, accuracy, and accessibility of EIS analysis, exhibiting great potential in accelerating the development of innovative electrochemical materials and devices.
Battery electrolytes are molecular mixtures that satisfy multiple objectives such as electrochemical stability and ion conductivity. In this work, we propose a differentiable chemical physics model for battery electrolytes, DiffMix, by incorporating geometric deep learning models (GDL) into a physics prior. We further developed an optimization algorithm and identified ionic conductivity peaks towards fast-charging battery applications. DiffMix expands the predictive modeling methods to provide low-cost methods for exploring chemical space of battery materials.
Harnessing the power of automation and machine learning, we're accelerating nanoparticle development for enhanced drug delivery.
The discovery and optimization of efficient and stable catalysts for the acidic oxygen evolution reaction (OER) are pivotal for advancing clean and sustainable energy sources. In our study, we offer a comprehensive analysis of our findings on the design and implementation of a self-driving laboratory for electrodeposited OER catalysts. We address the challenges posed by experimental variability and present a robust optimization framework. Our work provides valuable insights and considerations for future electrochemical self-driving laboratories.
The task of predicting whether a material can be made in a lab is difficult and depends on many factors. Previous works do not use thermodynamic stability, a relevant factor, when approaching this task. Thus, we tried using stability among other material descriptors to teach a machine learning model to predict synthesizability. Our model is accurate and identifies over 100 new materials as synthesizable.
PIGNet2: a "versatile" deep learning-based protein-ligand interaction prediction model comprising the potential in both lead optimization and virtual screening
Polymers are at the core of many advances in various fields: transportation, food packaging, energy storage, constructions. Current methods for selecting the proper materials rely on knowledge of polymer properties. Our objective with the Pyridine project is to deploy an artificial intelligence-based electronic interface that allows its user to input a set of targeted properties and the interface would output the optimal polymers to use, greatly accelerating R&D involving polymers.
Battery electrolytes are molecular mixtures that satisfy multiple objectives such as electrochemical stability and ion conductivity. In this work, we propose a differentiable chemical physics model for battery electrolytes, DiffMix, by incorporating geometric deep learning models (GDL) into a physics prior. We further developed an optimization algorithm and identified ionic conductivity peaks towards fast-charging battery applications. DiffMix expands the predictive modeling methods to provide low-cost methods for exploring chemical space of battery materials.
CrystaLLM is a large language model of the Crystallographic Information File (CIF) format that is capable of reliably generating correct CIF syntax, physically sensible parameters, and plausible crystal structures for many classes of inorganic compounds. It promises to be a useful tool for tasks requiring generative modelling of inorganic crystal structures.
Determining the structure of organic molecules is necessary for identifying unknown compounds and confirming synthesis products. Rotational spectroscopy provides rich but incomplete information about the 3D structure of molecules. We present two methods, a genetic algorithm, and an equivariant diffusion model, to predict the complete molecular structure starting from this incomplete information. Strong performance of the diffusion model points towards the feasibility of identifying molecules from just rotational spectroscopy and molecular formula.
Polyhydroxyalkanoates (PHAs) have drawn attention as sustainable replacements for petroleum-based plastics as the need for alternatives to these materials has increased. We represent Language models for PHA design (LaMP), fine-tuning state-of-the-art large language models to learn PHAs’ molecular structure-property relationship , and further effectively discover new PHAs based on its targeted properties for specific applications.
Self-driving labs (SDLs) powered by machine learning and automation are emerging technologies that can help overcome the time-consuming nature of electrochemical testing. By developing open-source tools for SDLs, we aim to make them accessible to researchers and educators, facilitating the rapid exploration and optimization of electroactive materials for various applications.
To enhance efficiency and reduce costs in chemistry labs, we propose a framework that integrates versatile robotic manipulators with various instruments commonly found in chemistry labs. Our approach successfully plans chemistry experiments using natural language, perceives transparent objects, simulates multi-stage manipulations, and executes robot tasks with safety considerations.
Graft injury affects over 30% of liver transplant recipients and the exact cause of injury can only be identified currently with a liver biopsy. Transplant recipients can develop graft injury due to various etiologies, such as T-cell mediated rejection (TCMR) and non-alcoholic steatohepatitis-LT (NASH-LT). We developed a machine learning tool, CleVER-LG, leveraging clinical variables and methylation patterns on circulating DNA in plasma as a non-invasive diagnostic tool of graft pathology.
Fused Deposition Modeling (FDM) is a popular and cost-effective 3D printing technology. However, calibrating FDM printers to new feedstocks can be a time-consuming trial and error process that acts as a barrier to new users. To address this, we developed a system for autonomous calibration of entry-level FDM printers using computer vision and metaheuristic methods. Results demonstrate that the system can autonomously calibrate printers to achieve submillimeter dimensional accuracy, surpassing the printer's published tolerance.
Optimizing 3D printing parameters for new materials is a very time consuming task, so we have developed an automated workflow involving pellet-based 3D printers, a collaborative robot, camera vision, and custom-designed lab equipment based on Arduino with the objective of automating all the parts of the experiment. In each iteration, we print and evaluate the quality of the material, and our Artificial Intelligence algorithm learns from the results and plans the next experiments.
Synthesis is the major bottleneck in the design of new inorganic materials. We developed a data-driven algorithm to rapidly design synthesis recipes that are faster, more energy efficient, and more selective than conventional approaches. When coupled with automated laboratory efforts, our approach promises to accelerate the time from lab to market for novel energy materials.
Making materials is often the bottleneck in (molecular) materials discovery – therefore, parallelizing synthesis can allow for significant acceleration. However, this necessitates an experiment planning AI that accounts for variable throughput or delayed experimental responses. Here, we describe an adaptive framework for asynchronous, parallel optimization, integrating multiple experiment “threads” at different locations. We demonstrate the ability to orchestrate a delocalized campaign towards organic solid-state lasing materials – showcasing the next step towards democratized materials discovery.
Our work developed an open-source tool that assists Electrochemical Impedance Spectroscopy (EIS) analysis, which is a crucial technique widely used in analyzing electrochemical materials. By automatically proposing statistically plausible circuit models, AutoEIS streamlines data interpretation without requiring in-depth knowledge of electrochemical systems. This generalized automated approach enhances the efficiency, accuracy, and accessibility of EIS analysis, exhibiting great potential in accelerating the development of innovative electrochemical materials and devices.
Harnessing the power of automation and machine learning, we're accelerating nanoparticle development for enhanced drug delivery.
We created a data management system to optimize the use of R&D data in hydrogen technology. The underlying data model is highly flexible in representing fabrication workflows, and uses ontologies to create highly connected knowledge graphs. The data structure allows the implementation of large-language models to increase the findability of data.
In this work we present the Workflow Execution Interface (WEI). Our presentation will explore how to optimize workflows, improve instrument integration, and facilitate the use of high-performance computing, thus laying the groundwork for autonomous discovery. We will present our progress integrating several different laboratories at Argonne National Laboratory and introduce the Rapid Prototyping Laboratory (RPL) as a student-friendly training ground for self-driving laboratories.
CrystaLLM is a large language model of the Crystallographic Information File (CIF) format that is capable of reliably generating correct CIF syntax, physically sensible parameters, and plausible crystal structures for many classes of inorganic compounds. It promises to be a useful tool for tasks requiring generative modelling of inorganic crystal structures.
Determining the structure of organic molecules is necessary for identifying unknown compounds and confirming synthesis products. Rotational spectroscopy provides rich but incomplete information about the 3D structure of molecules. We present two methods, a genetic algorithm, and an equivariant diffusion model, to predict the complete molecular structure starting from this incomplete information. Strong performance of the diffusion model points towards the feasibility of identifying molecules from just rotational spectroscopy and molecular formula.
Polyhydroxyalkanoates (PHAs) have drawn attention as sustainable replacements for petroleum-based plastics as the need for alternatives to these materials has increased. We represent Language models for PHA design (LaMP), fine-tuning state-of-the-art large language models to learn PHAs’ molecular structure-property relationship , and further effectively discover new PHAs based on its targeted properties for specific applications.
Self-driving labs (SDLs) powered by machine learning and automation are emerging technologies that can help overcome the time-consuming nature of electrochemical testing. By developing open-source tools for SDLs, we aim to make them accessible to researchers and educators, facilitating the rapid exploration and optimization of electroactive materials for various applications.
We created a data management system to optimize the use of R&D data in hydrogen technology. The underlying data model is highly flexible in representing fabrication workflows, and uses ontologies to create highly connected knowledge graphs. The data structure allows the implementation of large-language models to increase the findability of data.
To enhance efficiency and reduce costs in chemistry labs, we propose a framework that integrates versatile robotic manipulators with various instruments commonly found in chemistry labs. Our approach successfully plans chemistry experiments using natural language, perceives transparent objects, simulates multi-stage manipulations, and executes robot tasks with safety considerations.
The discovery and optimization of efficient and stable catalysts for the acidic oxygen evolution reaction (OER) are pivotal for advancing clean and sustainable energy sources. In our study, we offer a comprehensive analysis of our findings on the design and implementation of a self-driving laboratory for electrodeposited OER catalysts. We address the challenges posed by experimental variability and present a robust optimization framework. Our work provides valuable insights and considerations for future electrochemical self-driving laboratories.
Graft injury affects over 30% of liver transplant recipients and the exact cause of injury can only be identified currently with a liver biopsy. Transplant recipients can develop graft injury due to various etiologies, such as T-cell mediated rejection (TCMR) and non-alcoholic steatohepatitis-LT (NASH-LT). We developed a machine learning tool, CleVER-LG, leveraging clinical variables and methylation patterns on circulating DNA in plasma as a non-invasive diagnostic tool of graft pathology.
Fused Deposition Modeling (FDM) is a popular and cost-effective 3D printing technology. However, calibrating FDM printers to new feedstocks can be a time-consuming trial and error process that acts as a barrier to new users. To address this, we developed a system for autonomous calibration of entry-level FDM printers using computer vision and metaheuristic methods. Results demonstrate that the system can autonomously calibrate printers to achieve submillimeter dimensional accuracy, surpassing the printer's published tolerance.
Optimizing 3D printing parameters for new materials is a very time consuming task, so we have developed an automated workflow involving pellet-based 3D printers, a collaborative robot, camera vision, and custom-designed lab equipment based on Arduino with the objective of automating all the parts of the experiment. In each iteration, we print and evaluate the quality of the material, and our Artificial Intelligence algorithm learns from the results and plans the next experiments.
The task of predicting whether a material can be made in a lab is difficult and depends on many factors. Previous works do not use thermodynamic stability, a relevant factor, when approaching this task. Thus, we tried using stability among other material descriptors to teach a machine learning model to predict synthesizability. Our model is accurate and identifies over 100 new materials as synthesizable.
Synthesis is the major bottleneck in the design of new inorganic materials. We developed a data-driven algorithm to rapidly design synthesis recipes that are faster, more energy efficient, and more selective than conventional approaches. When coupled with automated laboratory efforts, our approach promises to accelerate the time from lab to market for novel energy materials.
PIGNet2: a "versatile" deep learning-based protein-ligand interaction prediction model comprising the potential in both lead optimization and virtual screening
Polymers are at the core of many advances in various fields: transportation, food packaging, energy storage, constructions. Current methods for selecting the proper materials rely on knowledge of polymer properties. Our objective with the Pyridine project is to deploy an artificial intelligence-based electronic interface that allows its user to input a set of targeted properties and the interface would output the optimal polymers to use, greatly accelerating R&D involving polymers.
Making materials is often the bottleneck in (molecular) materials discovery – therefore, parallelizing synthesis can allow for significant acceleration. However, this necessitates an experiment planning AI that accounts for variable throughput or delayed experimental responses. Here, we describe an adaptive framework for asynchronous, parallel optimization, integrating multiple experiment “threads” at different locations. We demonstrate the ability to orchestrate a delocalized campaign towards organic solid-state lasing materials – showcasing the next step towards democratized materials discovery.
In this work we present the Workflow Execution Interface (WEI). Our presentation will explore how to optimize workflows, improve instrument integration, and facilitate the use of high-performance computing, thus laying the groundwork for autonomous discovery. We will present our progress integrating several different laboratories at Argonne National Laboratory and introduce the Rapid Prototyping Laboratory (RPL) as a student-friendly training ground for self-driving laboratories.
Our work developed an open-source tool that assists Electrochemical Impedance Spectroscopy (EIS) analysis, which is a crucial technique widely used in analyzing electrochemical materials. By automatically proposing statistically plausible circuit models, AutoEIS streamlines data interpretation without requiring in-depth knowledge of electrochemical systems. This generalized automated approach enhances the efficiency, accuracy, and accessibility of EIS analysis, exhibiting great potential in accelerating the development of innovative electrochemical materials and devices.
Battery electrolytes are molecular mixtures that satisfy multiple objectives such as electrochemical stability and ion conductivity. In this work, we propose a differentiable chemical physics model for battery electrolytes, DiffMix, by incorporating geometric deep learning models (GDL) into a physics prior. We further developed an optimization algorithm and identified ionic conductivity peaks towards fast-charging battery applications. DiffMix expands the predictive modeling methods to provide low-cost methods for exploring chemical space of battery materials.
Harnessing the power of automation and machine learning, we're accelerating nanoparticle development for enhanced drug delivery.