materials-chemistry/ai produced the result/Frontiers in Chemistry 2023 · v2
Neural network potential used to simulate carbon dioxide moving inside porous frameworks
Researchers simulated carbon dioxide gas inside two porous crystalline materials to track how it is held and how it moves. A pre-trained neural network, ANI-2x, supplied the atomic energies and forces that drove every simulation.
spectrum · one line per step, placed by what the step does · bright lines used AI
Storage and diffusion of CO2 in covalent organic frameworks—A neural network-based molecular dynamics simulation approach
Frontiers in Chemistry, 2023
doi:10.3389/fchem.2023.1100210 · record aix-00094 v2 · checked 2026-10-08
- AI was for
- Simulation surrogate
- Model family
- Multilayer perceptron
- Checked by
- Replication
- Code
- not reported
The finding the paper is about came from the AI.
What this research was about
Covalent organic frameworks are crystalline solids built from light atoms joined by strong chemical bonds into a regular, open scaffold. The scaffold leaves channels and cavities, called pores, running through the material. Gas molecules can settle inside those pores, which is why such materials are studied as sponges for carbon dioxide. Knowing how much gas a framework will hold, and how quickly molecules can travel through its channels, means tracking the push and pull between the framework's atoms and the gas. Calculating those forces properly, from quantum mechanics, is slow. Doing it for thousands of atoms over millions of time steps, which is what a moving picture of the gas requires, is slower still.
The researchers studied two azine-linked frameworks: HEX-COF1, whose structure is built in flat sheets, and 3D-HNU5, whose scaffold extends in three directions. They set out to follow carbon dioxide inside each one across a range of gas loadings and temperatures, and to extract the pore sizes, the contacts between gas and framework, the speed of diffusion and an estimate of how much gas each material can take up.
Where AI came in
The gap between accuracy and speed was filled by a neural network potential: a network trained in advance on quantum chemical calculations, which then predicts the energy of an arrangement of atoms and the forces on each one. The researchers used an existing network, ANI-2x, as it came, without training it further. It supplied every energy and force in the molecular dynamics runs, so it stood in for the quantum chemical calculation that would otherwise have been done at each step. Before the main runs, they checked its figure for the interaction between a framework and a single carbon dioxide molecule against results from conventional quantum chemical methods.
Everything after the simulations was conventional analysis of the resulting trajectories, with no learning involved: the pore size distributions, the distance distributions that located the dominant contacts between framework hydrogen atoms and gas oxygen atoms, and the diffusion coefficients and activation energies. Because the trajectories themselves came from the network, those findings rest on it.
Written by AIxSci from the checked record below, to give context for readers outside the field. It is not part of the record.
The work
Technical · from the record
The ANI-2x neural network potential was used to supply the energies and forces for molecular dynamics simulations of CO2 inside two azine-linked covalent organic frameworks, the two-dimensional HEX-COF1 and the three-dimensional 3D-HNU5. Interaction energies between each framework and a single CO2 molecule were first compared with a set of DFT and DFTB methods, after which 3.5 ns sampling runs at a range of CO2 loadings and temperatures were analysed for structure, pore size distribution, minimum distance distribution functions and diffusion. The activation energy of diffusion reached its maximum at 24 CO2 molecules (15.6 w%) for 3D-HNU5 and at 10 CO2 molecules per pore (21.2 w%) for HEX-COF1, against reported experimental uptake capacities of 12.3 w% and 20.0 w%. The simulations place the dominant host-guest contacts between hydrogen atoms of the frameworks and oxygen atoms of CO2, and show a temperature- and loading-dependent structural deformation in 3D-HNU5 that is absent in HEX-COF1.
How AI was used
The pre-trained ANI-2x neural network potential was used without further training as the energy and force engine for periodic molecular dynamics simulations of two covalent organic frameworks with CO2 guests. Supercells were built from published unit cell data and lattice parameters, and ANI-2x was first used to optimise the pristine framework, an isolated CO2 molecule and the combined system so that its interaction energy could be set against DFT and DFTB references. Simulations then used the velocity-Verlet integrator with periodic boundary conditions, hydrogen-containing bonds constrained with SHAKE/RATTLE to a 2.0 fs time step, a Berendsen thermostat and a Monte-Carlo manostat, with empty frameworks equilibrated under NVT and then NPT before CO2 was introduced incrementally at a series of compositions and temperatures. The resulting trajectories were post-processed with conventional, non-learned analyses: minimum distance distribution functions decomposed into atomic contributions, geometric pore size distributions computed with PoreBlazer and Zeo++ over frames drawn from equilibrated runs, and diffusion coefficients from the Einstein relation fitted to the diffusive regime, from which activation energies of diffusion were obtained as a function of loading.
The shape of the work
Structural · the record, drawn
no AI
Construct COF simulation cells
Obtaining raw data, whether by measurement, download or retrieval.
the unit cell information reported by Guan and coworkers was employed as a starting pointwhere the paper describes this · verbatim
AI
Benchmark NNP host-gas interaction energy against DFT and DFTB
Testing outputs against ground truth. The AI stood in for simulation.
the structure of the pristine COF, a single CO2 molecule, and the CO2@COF system was optimized and the interaction energy calculatedwhere the paper describes this · verbatim
no AI
Load host structures with CO2 at set compositions
Cleaning, filtering, normalising or labelling data already obtained.
CO2 molecules were incrementally introduced into the pre-equilibrated host structurewhere the paper describes this · verbatim
AI
Run NNP molecular dynamics sampling
Numerical or physics simulation, including where a learned surrogate replaces it. The AI stood in for simulation.
Finally, 3.5 ns sampling runs under NPT conditions were conducted.where the paper describes this · verbatim
no AI
Minimum distance distribution function analysis
Extracting understanding from model behaviour.
minimum distance distribution functions (MDDFs) were calculatedwhere the paper describes this · verbatim
no AI
Compute pore size distributions from trajectory frames
Extracting understanding from model behaviour.
the associated pore size distributions (PSDs) for HEX-COF1 and 3D-HNU5 were calculated using two different geometric methodswhere the paper describes this · verbatim
no AI
Derive diffusion coefficients and activation energies
Extracting understanding from model behaviour.
the diffusion coefficient D of CO2 in accordance with the Einstein relation given in Eq. (2) was determined from the simulation trajectorieswhere the paper describes this · verbatim
no AI
Compare simulated uptake maxima with reported experimental capacities
Testing outputs against ground truth.
this data is in very good agreement with the empirically determined CO2 uptake capacity of 12.3 w% CO2 and 20.0 w% CO2where the paper describes this · verbatim
What the record says
Technical · every part carries its own basis
+ in the paper~ our reading− not reported
How to read the quotations. A quotation shows where the paper describes something. It does not quote every value beside it: one passage locates a part of the work, and values without their own quotation are our reading of that passage.
The neural network potential supplies all energies and forces for the molecular dynamics runs, so the reported diffusion coefficients, pore size distributions, distribution functions and uptake-capacity estimates are products of the model.
in conjunction with the neural network potential (NNP) ANI-2x employed to execute the energy and force calculationswhere the paper describes this · verbatim
this data is in very good agreement with the empirically determined CO2 uptake capacity of 12.3 w% CO2 and 20.0 w% CO2where the paper describes this · verbatim
The raw data supporting the conclusion of this article will be made available by the authors, without undue reservation.where the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- CodeWhether the code is available is not stated.
- Trained model weightsWhether the trained model is available is not stated.
- ComputeThe hardware or time used is not stated.
- How many were testedThe paper gives no count of what was tested.
- Version of ANI-2xWhich version of the model was used is not stated.
About this article
Record aix-00094, version 2, checked by a person on 2026-10-08. The record describes the paper; it does not assess whether the paper's findings are right. The paper is published under CC-BY; quotations are at most 25 words. How we work · Report an error