Periodic Reporting for period 4 - SubNano (Computational Photochemistry in the Long Timescale: Sub-ns Photoprocesses in DNA)
Okres sprawozdawczy: 2024-03-01 do 2024-08-31
Molecular excited states are at the core of a myriad of phenomena. They are crucial for Earth's biosphere, being responsible for the first step in harvesting sunlight through photosynthesis. They are equally vital for vision and bioluminescence. In technology, excited states are the basis of optoelectronics, from photovoltaics to light-emitting diodes.
Complementary to experiments, nonadiabatic molecular dynamics has been valuable in understanding these phenomena. It describes how the electronic density is distributed in an excited molecule and how it changes with time. Currently, available methods can predict the excited-state molecular evolution within very short timescales of about a few thousandth of nanoseconds. Nevertheless, many crucial processes happen on significantly longer scales.
SubNano aimed to create a series of computational methods to explore such long-timescale processes. The challenge was multiple: on the one hand, we needed to massively speed up our simulations; on the other hand, we needed to ensure that the quality of these simulations was not degraded by the long propagation times. Moreover, we needed ways to deal with the large amount of data these simulations should generate.
To face the speed-up challenge, we focused on surface hopping, a low-cost-dynamics method based on independent classical trajectories stochastically jumping between states. We rewrote the Newton-X program, developed by my group, from scratch to make it highly efficient. Then, we developed a new method to compute couplings at low computational costs. Finally, we developed machine learning algorithms to reduce computational costs.
On the accuracy front, the challenge was to deal with the deficient description of quantum features in surface hopping. These deficiencies tend to be negligible in the short time scale. However, they may accumulate to unacceptable levels over a long timescale. Thus, we developed methods to correct zero-point-energy leakages and quantum coherence inaccuracies.
To deal with the analysis of big data, we developed a dedicated program to perform advanced statistical analysis.
All methods developed during SubNano were implemented into programs, which were assembled into the open-source and freely available NX platform.
Performing nonadiabatic dynamics simulations in the long timescale simulation is still challenging. Nevertheless, the many methodological and computational developments we did in SubNano established the new state of the art in the field and started to make it an achievable goal.
Complementary to experiments, nonadiabatic molecular dynamics has been valuable in understanding these phenomena. It describes how the electronic density is distributed in an excited molecule and how it changes with time. Currently, available methods can predict the excited-state molecular evolution within very short timescales of about a few thousandth of nanoseconds. Nevertheless, many crucial processes happen on significantly longer scales.
SubNano aimed to create a series of computational methods to explore such long-timescale processes. The challenge was multiple: on the one hand, we needed to massively speed up our simulations; on the other hand, we needed to ensure that the quality of these simulations was not degraded by the long propagation times. Moreover, we needed ways to deal with the large amount of data these simulations should generate.
To face the speed-up challenge, we focused on surface hopping, a low-cost-dynamics method based on independent classical trajectories stochastically jumping between states. We rewrote the Newton-X program, developed by my group, from scratch to make it highly efficient. Then, we developed a new method to compute couplings at low computational costs. Finally, we developed machine learning algorithms to reduce computational costs.
On the accuracy front, the challenge was to deal with the deficient description of quantum features in surface hopping. These deficiencies tend to be negligible in the short time scale. However, they may accumulate to unacceptable levels over a long timescale. Thus, we developed methods to correct zero-point-energy leakages and quantum coherence inaccuracies.
To deal with the analysis of big data, we developed a dedicated program to perform advanced statistical analysis.
All methods developed during SubNano were implemented into programs, which were assembled into the open-source and freely available NX platform.
Performing nonadiabatic dynamics simulations in the long timescale simulation is still challenging. Nevertheless, the many methodological and computational developments we did in SubNano established the new state of the art in the field and started to make it an achievable goal.
We can group the SubNano tasks into five categories:
I. Methods development:
We developed and implemented several methods to improve the quality, broaden the application domain, and speed up the simulations.
We improved the quality with QDCT and LP-ZPE. Both methods fix quantum effects.
The MQC-PE method broadened the domain of application, allowing the simulation of dynamics induced by thermal light. TD-BA, a method to determine nonadiabatic couplings without wavefunction evaluation, also broadened the scope of dynamics.
We achieved a large simulation speed-up thanks to TD-BA and a series of machine-learning techniques.
II. Software implementation:
SubNano allowed the development of the NX platform. It started from the classical Newton-X program and gained a series of additions, including Newton-X NS, an entirely redesigned Newton-X version to improve performance, manage big data, and comply with open data standards; Ulamdyn, a program to analyze dynamics results based on unsupervised learning techniques; Legion, a program for trajectory-based quantum wavepacket dynamics; and Alexdyn, a program to control dynamics using socked-based interprocess communication. A new version of the PySOC for spin-orbit couplings was also developed. Moreover, we have contributed to implementing new versions of the MLatom for atomistic machine learning.
The programs developed by our groups are open-source and freely available for download.
III. Benchmarks construction:
During SubNano, we appraised several features impacting the simulations, including the performance of standard nonadiabatic dynamics methods in the long timescale, the effect of the electronic structure method on dynamics, and the conditions that should be satisfied and the algorithmic limitations of long-timescale dynamics, among others.
IV. Case studies with new dynamics protocols:
Among the several molecular systems investigated in SubNano, the following case studies deserve being mentioned, as they introduced innovative research protocols: Photoisomerization of trans-resveratrol with explore-then-assess protocol; Non-Kasha fluorescence of pyrene with ergotic protocol; Cytosine-solvent energy transfer modeling.
All data generated by the group is available in the form of public datasets.
V. Reviews, perspectives, and conceptual analyses:
During SubNano, we reviewed several aspects of the relevant literature, including mixed quantum-classical approaches, multireference methods, machine learning for excited-state simulations, and dynamics simulations in active environments.
SubNano also offered the opportunity to dive into conceptual questions underlying the nonadiabatic dynamics simulations, such as ultrafast internal conversion without energy crossing, criteria for classifying double excitations, and wavefunction collapse times.
I. Methods development:
We developed and implemented several methods to improve the quality, broaden the application domain, and speed up the simulations.
We improved the quality with QDCT and LP-ZPE. Both methods fix quantum effects.
The MQC-PE method broadened the domain of application, allowing the simulation of dynamics induced by thermal light. TD-BA, a method to determine nonadiabatic couplings without wavefunction evaluation, also broadened the scope of dynamics.
We achieved a large simulation speed-up thanks to TD-BA and a series of machine-learning techniques.
II. Software implementation:
SubNano allowed the development of the NX platform. It started from the classical Newton-X program and gained a series of additions, including Newton-X NS, an entirely redesigned Newton-X version to improve performance, manage big data, and comply with open data standards; Ulamdyn, a program to analyze dynamics results based on unsupervised learning techniques; Legion, a program for trajectory-based quantum wavepacket dynamics; and Alexdyn, a program to control dynamics using socked-based interprocess communication. A new version of the PySOC for spin-orbit couplings was also developed. Moreover, we have contributed to implementing new versions of the MLatom for atomistic machine learning.
The programs developed by our groups are open-source and freely available for download.
III. Benchmarks construction:
During SubNano, we appraised several features impacting the simulations, including the performance of standard nonadiabatic dynamics methods in the long timescale, the effect of the electronic structure method on dynamics, and the conditions that should be satisfied and the algorithmic limitations of long-timescale dynamics, among others.
IV. Case studies with new dynamics protocols:
Among the several molecular systems investigated in SubNano, the following case studies deserve being mentioned, as they introduced innovative research protocols: Photoisomerization of trans-resveratrol with explore-then-assess protocol; Non-Kasha fluorescence of pyrene with ergotic protocol; Cytosine-solvent energy transfer modeling.
All data generated by the group is available in the form of public datasets.
V. Reviews, perspectives, and conceptual analyses:
During SubNano, we reviewed several aspects of the relevant literature, including mixed quantum-classical approaches, multireference methods, machine learning for excited-state simulations, and dynamics simulations in active environments.
SubNano also offered the opportunity to dive into conceptual questions underlying the nonadiabatic dynamics simulations, such as ultrafast internal conversion without energy crossing, criteria for classifying double excitations, and wavefunction collapse times.
SubNano has defined the new state of the art of nonadiabatic dynamics on a long timescale. Among the main innovations, I can highlight:
* Development of Newton-X NS: this new version of the Newron-X program can be 1800 times faster than the classical Newton-X. Moreover, the ensemble of software innovation in SubNano gave rise to the NX platform. It includes Newton-X NS and several other programs that together enable simulations of nonadiabatic dynamics, from initial condition generation to advanced statistical analysis with unsupervised learning techniques.
* Development of QDCT method. Surface hopping is the only affordable method for long-timescale nonadiabatic dynamics. Nevertheless, it suffers from intrinsic handicaps in the description of quantum coherences. We solved this problem with QDCT, which integrates quantum equations of motion using surface hopping and interpolated trajectories as a grid.
* Development of the LP-ZPE correction. For decades, the ZPE leakage has been a problem haunting classical molecular dynamics simulations. We developed the LP-ZPE correction as a game-change: it corrects the ZPE leakage using a collisional model based on local vibrations, with minimum computational costs and interference in the dynamics.
* Development of TD-BA. One of the crucial quantities for propagating nonadiabatic dynamics is the nonadiabatic coupling. These quantities are not readily available for most quantum chemical methods. In SubNano, we introduced the time-dependent Baeck-An (TD-BA) method, which obtains the couplings from the curvature of trajectories. Because this quantity is always available during dynamics, we can now propagate dynamics with any method.
* Development of dynamics induced by thermal light with MQC-PE. Most simulations and experiments assume photoexcitation is triggered by coherent light. Nevertheless, natural photoprocesses induced by sunlight are triggered by incoherent light. In SubNano, I developed MQC-PE, a method to evaluate the effect of excitation by incoherent light in surface hopping.
* Development of Newton-X NS: this new version of the Newron-X program can be 1800 times faster than the classical Newton-X. Moreover, the ensemble of software innovation in SubNano gave rise to the NX platform. It includes Newton-X NS and several other programs that together enable simulations of nonadiabatic dynamics, from initial condition generation to advanced statistical analysis with unsupervised learning techniques.
* Development of QDCT method. Surface hopping is the only affordable method for long-timescale nonadiabatic dynamics. Nevertheless, it suffers from intrinsic handicaps in the description of quantum coherences. We solved this problem with QDCT, which integrates quantum equations of motion using surface hopping and interpolated trajectories as a grid.
* Development of the LP-ZPE correction. For decades, the ZPE leakage has been a problem haunting classical molecular dynamics simulations. We developed the LP-ZPE correction as a game-change: it corrects the ZPE leakage using a collisional model based on local vibrations, with minimum computational costs and interference in the dynamics.
* Development of TD-BA. One of the crucial quantities for propagating nonadiabatic dynamics is the nonadiabatic coupling. These quantities are not readily available for most quantum chemical methods. In SubNano, we introduced the time-dependent Baeck-An (TD-BA) method, which obtains the couplings from the curvature of trajectories. Because this quantity is always available during dynamics, we can now propagate dynamics with any method.
* Development of dynamics induced by thermal light with MQC-PE. Most simulations and experiments assume photoexcitation is triggered by coherent light. Nevertheless, natural photoprocesses induced by sunlight are triggered by incoherent light. In SubNano, I developed MQC-PE, a method to evaluate the effect of excitation by incoherent light in surface hopping.