Patent Application
20250166741
This application is based upon and claims priority to Chinese Patent Application No. 202311547565.7, filed on Nov. 20, 2023, the entire contents of which are incorporated herein by reference.
The present disclosure relates to the prediction of a hydration structure of an acid radical anion, and in particular to a machine learning (ML)-accelerated first-principles prediction method for a hydration structure of an acid radical anion.
The hydration structure around ions has a significant impact on the electron transfer and reaction rate of chemical reaction processes. Hydration can change the driving force of chemical reactions, making it a key factor to consider in selective actions. Currently, there are reports and studies on the hydration structures of metal cations, such as Pb(II), Cu(II), Fe(II), Fe(III), Zn(II), V(II), V(III), Ca(II), and Mg(II), in aqueous solutions. However, there are few studies on the hydration structures of acid radical anions formed by metal cations and oxygen or sulfur, such as WO42− and MoS42−.
In order to more accurately simulate the real environment of ion hydration, it is necessary to consider the large number of water molecules around the ions, which makes it impossible for precise quantitative measurements in traditional experiments. Quantum chemistry theory calculations can simulate and predict complex chemical reactions through high-speed computers and precise quantum mechanics algorithms. At present, ab initio molecular dynamics (AIMD) simulations have basically solved the problem of “inaccuracy”, but the problems of “slow calculation” and “high cost” still exist. For example, the simulation of a WO42−_ 100H2O_2Na system includes 4,000 steps, taking a total of 2 ps, 0.5 fs of time for each step, conducted on a high-performance computing server with a total of 96 cores at 2 nodes. It requires uninterrupted operation for about 12 days, resulting in a computing service fee of approximately 1,400 yuan (0.05 yuan/core hour). Obviously, a 100 ps AIMD simulation requires uninterrupted operation for about 60 days and costs approximately 70,000 yuan. Such slow computing progress and expensive computing costs cannot achieve the cost-effective purpose.
Therefore, it is necessary to establish a systematic research method to accurately and quickly predict the hydration structures of ions, especially acid radical anions such as WO42− and MoS42−.
In view of the problems of “slow calculation” and “high cost” in predicting the hydration structures of acid radical anions, the present disclosure provides a machine learning (ML)-accelerated first-principles prediction method for a hydration structure of an acid radical anion. The prediction method greatly improves computational efficiency and reduces computational costs while ensuring computational accuracy.
In order to achieve the above objective, the present disclosure provides a ML-accelerated first-principles prediction method for a hydration structure of an acid radical anion (such as WO42−, MoS42−, SO42−, or CO32−), including the following steps:
The above technical solution achieves a deep potential molecular dynamics (DPMD) simulation on the anion hydration structure, especially achieves the first rapid and accurate prediction for the hydration structure of the acid radical anion. A traditional AIMD simulation method can generally only achieve a 1-10 ps simulation, while the DPMD simulation method can easily achieve a 100-1,000 ps simulation. The present disclosure greatly improves computational efficiency and reduces computational costs while ensuring computational accuracy.
Specifically, in the step S1, a method for constructing the anion hydration structure through Materials Studio software includes: creating a periodic simulation box, placing the acid radical anion M at a center of the box, uniformly and randomly distributing m H2O around the acid radical anion M, and placing an alkali metal cation at an edge of the box to balance a charge, thereby completing the construction of the initial anion hydration structure M_mH2O, where m is above 40.
In the above technical solution, careful consideration should be given when selecting the initial arrangement of H2O and the initial distance between M and H2O to ensure that the simulated initial state is reasonable.
On the basis of the above technical solution, in order to more accurately describe and construct the hydration structure of the anion M, before constructing the anion hydration structure through first-principles density functional theory (DFT), the prediction method further includes pre-optimizing: preparing a def2-TZVP, def2-SVP, def2-QZVP, def2-TZVPP, cc-pVDZ-PP or aug-cc-pVDZ-PP basis set through an ωB97XD, B3LYP or PBE functional; and pre-optimizing a simple hydration structure M_nH2O formed by n H2O in a sternest layer around the anion M to acquire a stern layer hydration structure of the acid radical anion M. The step S1 further includes: uniformly and randomly distributing remaining (m-n) H2O around M_nH2O, thereby completing the construction of the initial anion hydration structure M_mH2O; and optimizing the constructed anion hydration structure through Vienna ab initio simulation package (VASP).
Specifically, in the step S2, the perturbing includes: changing an atomic coordinate position and a size of the simulation box to generate A different perturbed structures, each perturbed structure being in a canonical ensemble (NVT ensemble) at 298.15 K; and recording B frames for each configuration at a time step of 0.5-1.5 fs, and conducting a short-term (step size×frame rate) AIMD simulation to generate a training dataset of A×B frames as basic training data for a DeepMD model, where A is 15-30; and B is 15-25.
Preferably, the time step is 0.5 fs; A is 20; and B is 20.
The perturbation is to appropriately change the atomic coordinate position and the size of the simulation box based on the already optimized structure, in order to further acquire more acceptable structures and generate the training dataset.
Specifically, the step S3 includes: conducting the ML force field training on the training dataset through deep potential generator (DP-GEN) and DeePMD kit software package, where the ML force field training includes 2×105 to 8×105 steps, with 2-5 independent ML models established each time. These models use the same reference dataset but have different initial weight values. Preferably, there are 4×105 training steps and 4 independent ML models.
Specifically, in the step S4, the conducting a molecular dynamics simulation includes: conducting, by a large-scale atomistic/molecular massively parallel simulator (LAMMPS), the molecular dynamics simulation on the anion hydration structure in the NVT ensemble under at least five different temperature conditions; and identifying, during the simulation, atomic structures with a force deviation ranging within 0.11-0.30 eV/Å as candidate configurations, where a maximum of 300 candidate configurations are selected.
The LAMMPS interface is complete and easy to expand, allowing for fast molecular dynamics parallel simulations on force field models trained through ML force field trainings, in order to select candidate configurations within the range of force deviations.
In a preferred implementation of the present disclosure, the prediction method includes: conducting, by the LAMMPS, the molecular dynamics simulation on the anion hydration structure in the NVT ensemble under five different temperature conditions of 250 K, 280 K, 300 K, 320 K, and 350 K.
For the selection of these five temperatures, a temperature gradient is constructed within the range of 250-350 K, taking into account common temperature conditions to ensure a more reasonable hydration structure.
Specifically, the step S5 includes: merging a candidate configuration with an energy and atomic force validated through the DFT into the training set for the subsequent iteration to further refine and train the ML model. For all iterative processes, automation is achieved through the DP-GEN software package until the model converges to acquire an accurate DP model.
Specifically, the step S6 includes: applying a trained ML force field to the LAMMPS, and conducting a 50-1,000 ps ML-accelerated deep potential molecular dynamics (i.e. DPMD) simulation on the DP model to ultimately acquire the hydration structure of the acid radical anion.
The present disclosure adopts the DPMD method, which greatly improves the simulation calculation speed and enables the accuracy of simulation results to be comparable to that of traditional AIMD methods, providing an efficient and accurate ML-accelerated method for first-principles prediction of anion hydration structures. Compared with traditional AIMD methods, in the analysis and validation stage of simulation results, the present disclosure can further validate the accuracy and efficiency of the ML-accelerated method, while exploring the impact of different simulation durations and conditions on simulation results and how to optimize and improve the technical solution to enhance accuracy and efficiency.
With the above technical solution, the present disclosure achieves the following beneficial effects:
The present disclosure implements the construction of the anion hydration structure, perturbation, ML force field training, molecular dynamics simulation, and model convergence. The present disclosure ultimately achieves DPMD simulation on the anion hydration structure, particularly achieving the first rapid and accurate prediction for the hydration structure of the acid radical anion. The traditional AIMD simulation method can generally only achieve a 1-10 ps simulation, while the DPMD simulation method can easily achieve a 100-1,000 ps simulation. The present disclosure greatly improves computational efficiency and reduces computational costs while ensuring computational accuracy.
The specific implementation of the present disclosure is described in detail below based on embodiments. It should be understood that the specific implementations described herein are merely intended to illustrate and interpret the present disclosure, rather than to limit the present disclosure.
As shown in
In order to more accurately describe and construct a hydration structure of MoS42−, based on first-principles DFT, the structure of MoS42−_6H2O is pre-optimized through a def2-TZVP basis set prepared by an ωB97XD functional, thereby generating a stern layer hydration structure of MoS42−.
The anion hydration structure is constructed through Materials Studio software. As shown in
The optimized MoS42−_64H2O_2Na structure is perturbed, including changing an atomic coordinate position and a size of the simulation box to generate 20 different perturbed structures. A short-term AIMD simulation is conducted on the optimized structure MoS42−_64H2O_2Na in an NVT ensemble at 298.15 K, with each configuration spanning 20 frames based on a time step of 0.5 fs.
In this way, a training dataset of the hydration structure MoS42−_64H2O_2Na is acquired as basic training data for a DeepMD model. It includes 20 perturbed structures, each with 20 frames, totaling 400 frames.
Through DP-GEN and DeePMD kit software package, a ML force fields training is conducted on the training dataset of the hydration structure MoS42−_64H2O_2Na.
Firstly, a 4×105-step training process is conducted on the training dataset of MoS42−_64H2O_2Na. 4 independent ML models are established for each training session. These models use the same reference dataset but have different initial weight values.
Then, the trained ML model is applied to a molecular dynamics simulation to explore a new anion hydration structure. The molecular dynamics simulation is conducted by a LAMMPS on the anion hydration structure MoS42−_64H2O_2Na in the NVT ensemble under five different temperature conditions of 250 K, 280 K, 300 K, 320 K, and 350 K. During the simulation, atomic structures with a force deviation ranging within 0.11-0.30 eV/Å are identified as candidate configurations, and a maximum of 300 candidate configurations are selected.
Finally, a candidate configuration with an energy and atomic force validated through the DFT is merged into the training set for the subsequent iteration to further refine and train the ML model. All iterative processes are automated through the DP-GEN software package until the model converges (i.e., the energy accuracy of the sampled data reaches over 95%) to acquire an accurate DP model. During the validation and testing phase of the model, different evaluation metrics and validation datasets are used to ensure the accuracy and generalization ability of the model.
A trained ML force field is applied to the LAMMPS, and a 100 ps ML-accelerated molecular dynamics (i.e. DPMD) simulation is conducted on the DP model, ultimately acquiring the hydration structure of MoS42−_64H2O_2Na (
The computational costs are as follows.
In terms of pre-optimization of MoS42−_6H2O, running continuously for 2 hours on a high-performance computing server with a total of 48 cores at one node results in a computing service fee of approximately 4.8 yuan (0.05 yuan/core hour).
In terms of generation of the MoS42−_64H2O_2Na dataset, running continuously for 2 hours on a high-performance computing server with a total of 96 cores at 2 nodes results in a computing service fee of approximately 9.6 yuan (0.05 yuan/core hour), and 20 perturbed structures take a total of 40 hours, resulting in a computing service fee of 192 yuan.
In terms of training of the MoS42−_64H2O_2Na model, a round of iteration of the dataset on a 3090Ti graphics card takes 7 days. To ensure the accuracy of the model, 3 rounds of iterations are conducted, which take a total of 504 hours, resulting in a computational service fee of approximately 1,028.16 yuan (2.04 yuan/core hour).
In terms of the DPDM simulation of MoS42−_64H2O_2Na, a 100 ps trajectory generated by LAMMPS takes 2 hours on the 3090Ti graphics card, resulting in a computational service fee of approximately 4.08 yuan (2.04 yuan/core hour).
In this way, the hydration structure MoS42−_64H2O_2Na is acquired through the systematic 100 ps DPDM simulation, which takes 548 hours and a total of 1,229.04 yuan. In contrast, a 100 ps AIMD simulation requires uninterrupted operation for about 60 days and costs about 70,000 yuan. Therefore, the ML-accelerated anion hydration structure calculation method proposed by the present disclosure improves computational efficiency and reduces computational costs while ensuring computational accuracy.
As shown in
In order to more accurately describe and construct a hydration structure of WO42−, based on first-principles DFT, the structure of WO42−_6H2O is optimized through a def2-TZVP basis set prepared by an ωB97XD functional, thereby generating a stern layer hydration structure of WO42−.
The anion hydration structure is constructed through Materials Studio software. As shown in
The optimized WO42−_64H2O_2Na structure is perturbed, including changing an atomic coordinate position and a size of the simulation box to generate 20 different perturbed structures. A short-term AIMD simulation is conducted on the optimized structure WO42−_64H2O_2Na in an NVT ensemble at 298.15 K, with each configuration spanning 20 frames based on a time step of 0.5 fs.
In this way, a training dataset of the hydration structure WO42−_64H2O_2Na is acquired as basic training data for a DeepMD model. It includes 20 perturbed structures, each with 20 frames, totaling 400 frames.
Through DP-GEN and DeePMD kit software package, a ML force fields training is conducted on the training dataset of the hydration structure WO42−_64H2O_2Na.
Firstly, a 4×105-step training process is conducted on the training dataset of WO42−_64H2O_2Na. 4 independent ML models are established for each training session. These models use the same reference dataset but have different initial weight values.
Then, the trained ML model is applied to a molecular dynamics simulation to explore a new anion hydration structure. The molecular dynamics simulation is conducted by a LAMMPS on the anion hydration structure WO42−_64H2O_2Na in the NVT ensemble under five different temperature conditions of 250 K, 280 K, 300 K, 320 K, and 350 K. During the simulation, atomic structures with a force deviation ranging within 0.11-0.30 eV/Å are identified as candidate configurations, and a maximum of 300 candidate configurations are selected.
Finally, a candidate configuration with an energy and atomic force validated through the DFT is merged into the training set for the subsequent iteration to further refine and train the ML model. All iterative processes are automated through the DP-GEN software package until the model converges (i.e., the energy accuracy of the sampled data reaches over 95%) to acquire an accurate DP model. During the validation and testing phase of the model, different evaluation metrics and validation datasets are used to ensure the accuracy and generalization ability of the model.
A trained ML force field is applied to the LAMMPS, and a 100 ps ML-accelerated molecular dynamics (i.e. DPMD) simulation is conducted on the DP model, ultimately acquiring the hydration structure of WO42−_64H2O_2Na (
The computational costs are as follows.
In terms of pre-optimization of WO42−_6H2O, running continuously for 2 hours on a high-performance computing server with a total of 48 cores at one node results in a computing service fee of approximately 4.8 yuan (0.05 yuan/core hour).
In terms of generation of the WO42−_64H2O_2Na dataset, running continuously for 2 hours on a high-performance computing server with a total of 96 cores at 2 nodes results in a computing service fee of approximately 9.6 yuan (0.05 yuan/core hour), and 20 perturbed structures take a total of 40 hours, resulting in a computing service fee of 192 yuan.
In terms of training of the WO42−_64H2O_2Na model, a round of iteration of the dataset on a 3090Ti graphics card takes 7 days. To ensure the accuracy of the model, 3 rounds of iterations are conducted, which take a total of 504 hours, resulting in a computational service fee of approximately 1,028.16 yuan (2.04 yuan/core hour).
In terms of the DPDM simulation of WO42−_64H2O_2Na, a 100 ps trajectory generated by LAMMPS takes 2 hours on the 3090Ti graphics card, resulting in a computational service fee of approximately 4.08 yuan (2.04 yuan/core hour).
In this way, the hydration structure WO42−_64H2O_2Na is acquired through the systematic 100 ps DPDM simulation, which takes 548 hours and a total of 1,229.04 yuan. In contrast, a 100 ps AIMD simulation requires uninterrupted operation for about 60 days and costs about 70,000 yuan. Therefore, the ML-accelerated anion hydration structure calculation method proposed by the present disclosure improves computational efficiency and reduces computational costs while ensuring computational accuracy.
In the above embodiments, there are differences in the hydration structures of WO42− and MoS42− anions, with WO42− having a thicker hydration layer and a larger hydration coordination number than MoS42−. Therefore, compared to MoS42−, WO42− undergoes stronger hydration shielding in the aqueous solution, resulting in selective separation of WO42− and MoS42− by quaternary ammonium cations during extraction or ion exchange resin adsorption.
The hydration structures of SO42− and CO32− are predicted according to the method in Embodiment 1, and the hydration structures of [SO4(H2O)64]2− and [CO3(H2O)64]2− in the 100 ps DPMD simulation are acquired. As shown in
In this way, the hydration structures of SO42−_64H2O_2Na and CO32−_64H2O_2Na are acquired through the systematic 100 ps DPDM simulation, each taking approximately 500 hours and approximately 1,120 yuan. In contrast, a 100 ps AIMD simulation requires uninterrupted operation for about 60 days and costs about 70,000 yuan. Therefore, the ML-accelerated anion hydration structure calculation method proposed by the present disclosure improves computational efficiency and reduces computational costs while ensuring computational accuracy.
The above prediction method of the present disclosure can quickly and accurately predict the hydration behaviors of acid radical anions (such as WO42−, MoS42−, SO42−, CO32−) under different solution conditions, has broad application prospects in complex hydrochemistry, electrochemistry, environmental science, biochemistry, and other fields, and can accelerate related scientific research and technological development processes. For example, the method can be used to simulate and predict the migration and transformation processes of pollutants in water bodies, which can help develop pollution control strategies and environmental remediation plans. The method can help understand the hydration behaviors of acid radical anions in drug molecules, and optimize drug solubility and stability research. In addition, the method can be used to study anion binding sites in biomacromolecule (such as proteins and nucleic acids), and understand their functions in biological systems.
The above embodiments describe the preferred implementations of the present disclosure in detail. However, the present disclosure is not limited to specific details of the above implementations. A plurality of simple variations can be made to the technical solutions of the present disclosure within the scope of the technical idea of the present disclosure, and all of these simple variations fall within the protection scope of the present disclosure.
In addition, it should be noted that various specific technical features described in the above specific embodiments can be combined in any suitable manner, provided that there is no contradiction. To avoid unnecessary repetition, various possible combination modes of the present disclosure are not described separately.
In addition, different implementations of the present disclosure can also be combined arbitrarily. The combinations should also be regarded as the content disclosed in the present disclosure, provided that they do not violate the ideas of the present disclosure.
| Number | Date | Country | Kind |
|---|---|---|---|
| 202311547565 .7 | Nov 2023 | CN | national |
