Periodic Reporting for period 3 - ADAM (Autonomous Discovery of Advanced Materials)
Reporting period: 2023-10-01 to 2025-03-31
Materials impact most aspects of our lives, including healthcare, energy production, data storage and pollution control. However, the design of functional materials cannot be approached with the certainty and the engineering rules that would be used in planning and constructing a macroscopic object, such as a car or bridge. To be experimentally realisable, candidate materials must be reasonably stable with respect to changes in the crystal packing; to be functionally promising, they must exhibit favourable properties. The stability of a given hypothetical configuration is difficult to predict, and the “possibility space” of different geometric arrangements and chemical compositions – both of which impact the resulting properties -- is enormous and challenging to explore. Even once promising candidates are proposed, the synthesis and characterisation of such materials is not straightforward, potentially requiring a great deal of trial-and-error and exhaustive surveys of synthetic routes and conditions.
Our project aims to change the way that we discover new molecular materials by revolutionizing the exploration process. The programme integrates state-of-the-art computational chemistry methods, synthetic chemistry expertise, and laboratory robotics to transform our materials discovery capabilities. Our vision is of two related scientific “engines”. Firstly, a Computational Engine will perform intelligent, evolutionary exploration of the possibility space, using organic crystal structure prediction (CSP) methods and artificial intelligence (AI, or “machine learning”, ML) to explore the relationships between chemical composition, stable crystal structures, and desirable properties. Secondly, an Experimental Engine will carry out autonomous synthesis, characterisation, and properties testing of proposed materials using cutting-edge, AI-enhanced mobile “robot chemists”. The ultimate vision of ADAM is to couple these two engines together, creating an autonomous discovery platform that amplifies human creativity by searching the vast, unexplored chemical space for new materials with step-change properties.
Our project aims to change the way that we discover new molecular materials by revolutionizing the exploration process. The programme integrates state-of-the-art computational chemistry methods, synthetic chemistry expertise, and laboratory robotics to transform our materials discovery capabilities. Our vision is of two related scientific “engines”. Firstly, a Computational Engine will perform intelligent, evolutionary exploration of the possibility space, using organic crystal structure prediction (CSP) methods and artificial intelligence (AI, or “machine learning”, ML) to explore the relationships between chemical composition, stable crystal structures, and desirable properties. Secondly, an Experimental Engine will carry out autonomous synthesis, characterisation, and properties testing of proposed materials using cutting-edge, AI-enhanced mobile “robot chemists”. The ultimate vision of ADAM is to couple these two engines together, creating an autonomous discovery platform that amplifies human creativity by searching the vast, unexplored chemical space for new materials with step-change properties.
1st Reporting Period: Oct 2020 - Sept 2022
Development of our Computational Engine has proceeded along two strands. The first is the improvement and application of our methods for carrying out CSP for flexible molecules (whose ability to adopt multiple different molecular geometries increases the complexity of the possibility space); the primary result thus far is the streamlining and generalising of our in-house methods for performing such calculations. The second strand is in the development of our evolutionary chemical space exploration algorithm, and in particular the incorporation of predicted crystalline properties (via CSP) as a component of our algorithm’s measure of “fitness”. Having implemented this, our search algorithm is now capable of targeting regions of chemical space that lead to crystal structures with specified bulk properties, rather than relying solely on molecular information.
Regarding the Experimental Engine, much of our initial efforts have been focussed on automating characterisation and sample preparation processes, as these were previously the tasks that required the most human intervention and attention. As a result of these, we have made very significant progress with automated preparation of crystal samples for powder X-ray diffraction measurements (a form of structure elucidation), as well as the automated, AI-assisted monitoring of both solubility experiments and crystallisation attempts. We have additionally employed ML methods to train a robotic setup to manipulate standard lab glassware, as well as developing a small, mobile robot-mounted volatile organic compound sensor to improve the safety of our automated workflow.
2nd Reporting Period: Oct 2022 - Sept 2024
The project has continued to develop both the Computational and Experimental Engines described in the project brief. The Computational Engine has demonstrated remarkable scalability and generality in this period, performing CSP on an unprecedented scale for over 1,000 molecules, as well as guiding the synthesis of new porous materials. Our evolutionary algorithm for chemical exploration is also advancing, with new metrics for fitness and accelerated, multi-objective optimisation implemented. We are also designing and implementing a central database of CSP results suitable for programmatic access, to serve as the foundation of data exchange with the Experimental Engine.
Regarding the Experimental Engine, our static robot arm system for automated crystallisation and powder X-ray diffraction sample preparation is maturing and has demonstrated enormous potential for the autonomous crystallisation and characterisation of materials. We have additionally implemented an all-in-one in-fumehood robotics setup for synthesis and characterisation named "Robinhood", and demonstrated the ability of twinned mobile robotics setups to carry out synthesis and characterisation workflows in parallel. We are additionally expanding our toolkit of internet-of-things modules for monitoring the laboratory environment to ensure safety and to permit reliable navigation by mobile robots. We have demonstrated the power of these and other experimental approaches to discover multiple new porous materials with remarkable properties, particularly when guided by our Computational Engine.
Development of our Computational Engine has proceeded along two strands. The first is the improvement and application of our methods for carrying out CSP for flexible molecules (whose ability to adopt multiple different molecular geometries increases the complexity of the possibility space); the primary result thus far is the streamlining and generalising of our in-house methods for performing such calculations. The second strand is in the development of our evolutionary chemical space exploration algorithm, and in particular the incorporation of predicted crystalline properties (via CSP) as a component of our algorithm’s measure of “fitness”. Having implemented this, our search algorithm is now capable of targeting regions of chemical space that lead to crystal structures with specified bulk properties, rather than relying solely on molecular information.
Regarding the Experimental Engine, much of our initial efforts have been focussed on automating characterisation and sample preparation processes, as these were previously the tasks that required the most human intervention and attention. As a result of these, we have made very significant progress with automated preparation of crystal samples for powder X-ray diffraction measurements (a form of structure elucidation), as well as the automated, AI-assisted monitoring of both solubility experiments and crystallisation attempts. We have additionally employed ML methods to train a robotic setup to manipulate standard lab glassware, as well as developing a small, mobile robot-mounted volatile organic compound sensor to improve the safety of our automated workflow.
2nd Reporting Period: Oct 2022 - Sept 2024
The project has continued to develop both the Computational and Experimental Engines described in the project brief. The Computational Engine has demonstrated remarkable scalability and generality in this period, performing CSP on an unprecedented scale for over 1,000 molecules, as well as guiding the synthesis of new porous materials. Our evolutionary algorithm for chemical exploration is also advancing, with new metrics for fitness and accelerated, multi-objective optimisation implemented. We are also designing and implementing a central database of CSP results suitable for programmatic access, to serve as the foundation of data exchange with the Experimental Engine.
Regarding the Experimental Engine, our static robot arm system for automated crystallisation and powder X-ray diffraction sample preparation is maturing and has demonstrated enormous potential for the autonomous crystallisation and characterisation of materials. We have additionally implemented an all-in-one in-fumehood robotics setup for synthesis and characterisation named "Robinhood", and demonstrated the ability of twinned mobile robotics setups to carry out synthesis and characterisation workflows in parallel. We are additionally expanding our toolkit of internet-of-things modules for monitoring the laboratory environment to ensure safety and to permit reliable navigation by mobile robots. We have demonstrated the power of these and other experimental approaches to discover multiple new porous materials with remarkable properties, particularly when guided by our Computational Engine.
1st Reporting Period: Oct 2020 - Sept 2022
Two of our major achievements thus far – the incorporation of CSP-derived properties into measuring fitness our evolutionary algorithm, and our approach to training robots to manipulate glassware – represent, to our knowledge, novel achievements in their specific contexts. Our solubility screening workflow, assisted by AI, additionally represents a significant breakthrough in autonomous monitoring of characterisation experiments.
In the next reporting period, we hope to demonstrate progress in several subprojects: the use of Bayesian Optimisation approaches for managing autonomous experimental exploration of chemical space, the use of our evolutionary approach to direct automated synthesis of a novel materials system exhibiting a targeted property (initially targeting porous materials), and to begin the coupling of our Computational and Experimental Engines to allow direct, automatic feedback between the two.
2nd Reporting Period: Oct 2022 - Sept 2024
Our static robot setup represents a step-change in the automation of crystallisation experiments and preparation of samples for characterisation. Our collaborative CSP and experimental screening work for non-metal organic frameworks has unlocked entirely new strategies for the design and synthesis of porous materials and has been recognised with publication in Nature. Our upcoming combinatorial hydrogen-bonded organic framework survey will demonstrate the power of CSP to screen potential molecules for those most likely to lead to desirable crystal structures, which once prioritised for automated synthesis and characterisation can create novel materials from unexpected candidate species. Finally, our 1,000-molecule CSP -- to our knowledge the largest ever performed -- proved the scalabilty and reliability of the core of our Computational Engine and has unlocked the possibility of training new ML methods on unprecedented quantities of reliable data, expediting and strengthening the impact of our calculations.
In the next reporting period, we hope to publish the results of our first cross-site automated crystallisation and measurement experiments to prove the reliability and consistency of our setup, to serve as a proof-of-concept of coupling our two Engines, and to involve team members across all three partner sites in one collaborative study. We aim to expand upon our automated synthesis and characterisation setups to consider new classes of materials, guided by larger and more efficient CSP calculations than ever before. Finally, we intend to solidfy the foundations of the interface between our two Engines, in pursuit of our ultimate aim of a closed-loop materials discovery process.
Two of our major achievements thus far – the incorporation of CSP-derived properties into measuring fitness our evolutionary algorithm, and our approach to training robots to manipulate glassware – represent, to our knowledge, novel achievements in their specific contexts. Our solubility screening workflow, assisted by AI, additionally represents a significant breakthrough in autonomous monitoring of characterisation experiments.
In the next reporting period, we hope to demonstrate progress in several subprojects: the use of Bayesian Optimisation approaches for managing autonomous experimental exploration of chemical space, the use of our evolutionary approach to direct automated synthesis of a novel materials system exhibiting a targeted property (initially targeting porous materials), and to begin the coupling of our Computational and Experimental Engines to allow direct, automatic feedback between the two.
2nd Reporting Period: Oct 2022 - Sept 2024
Our static robot setup represents a step-change in the automation of crystallisation experiments and preparation of samples for characterisation. Our collaborative CSP and experimental screening work for non-metal organic frameworks has unlocked entirely new strategies for the design and synthesis of porous materials and has been recognised with publication in Nature. Our upcoming combinatorial hydrogen-bonded organic framework survey will demonstrate the power of CSP to screen potential molecules for those most likely to lead to desirable crystal structures, which once prioritised for automated synthesis and characterisation can create novel materials from unexpected candidate species. Finally, our 1,000-molecule CSP -- to our knowledge the largest ever performed -- proved the scalabilty and reliability of the core of our Computational Engine and has unlocked the possibility of training new ML methods on unprecedented quantities of reliable data, expediting and strengthening the impact of our calculations.
In the next reporting period, we hope to publish the results of our first cross-site automated crystallisation and measurement experiments to prove the reliability and consistency of our setup, to serve as a proof-of-concept of coupling our two Engines, and to involve team members across all three partner sites in one collaborative study. We aim to expand upon our automated synthesis and characterisation setups to consider new classes of materials, guided by larger and more efficient CSP calculations than ever before. Finally, we intend to solidfy the foundations of the interface between our two Engines, in pursuit of our ultimate aim of a closed-loop materials discovery process.