Supplementary data: Accurate large-scale simulations of siliceous zeolites by neural network potentials

Content

1. Zeolite databases

Deem database containing 331170 hypothetical zeolite frameworks [Deem09, Pophale11] geometrically optimized at the NNPscan level (note, the first row of the database is alpha-quartz): "DEEM_NNPscan.db"

Database of 236 exiting zeolite frameworks of the International Zeolite Association (IZA) optimized at the NNPscan level: "IZA_NNPscan.db"

Both databases are ASE SQLite database files of the Atomic Simulation Environment containing the ASE Atoms objects with energies and forces (NNPscan level); readable with ASE's I/O module

Additionally, relevant quantities can be extracted with, e.g., the following queries (further information: ase db --help):

ase db DEEM_NNPscan.db -c id,formula,natoms,volume,mass,density,energy_per_tsite,n_tsites,relative_energy
# Output
id|formula|natoms| volume| mass|density|energy_per_tsite|n_tsites|relative_energy
1|O6Si3 | 9|111.161|180.249| 26.988| -31.796| 3| 0.000
2|O16Si8 | 24|433.858|480.664| 18.439| -31.638| 8| 15.265
3|O16Si8 | 24|421.114|480.664| 18.997| -31.596| 8| 19.359
4|O16Si8 | 24|426.557|480.664| 18.755| -31.614| 8| 17.613
5|O16Si8 | 24|412.410|480.664| 19.398| -31.613| 8| 17.677
6|O16Si8 | 24|393.544|480.664| 20.328| -31.594| 8| 19.546
7|O16Si8 | 24|422.400|480.664| 18.939| -31.657| 8| 13.476
8|O16Si8 | 24|394.405|480.664| 20.284| -31.581| 8| 20.797
9|O12Si6 | 18|265.201|360.498| 22.624| -31.611| 6| 17.868
10|O16Si8 | 24|357.047|480.664| 22.406| -31.581| 8| 20.785
11|O16Si8 | 24|434.894|480.664| 18.395| -31.621| 8| 16.911
12|O16Si8 | 24|384.158|480.664| 20.825| -31.657| 8| 13.448
13|O12Si6 | 18|258.977|360.498| 23.168| -31.679| 6| 11.278
14|O16Si8 | 24|466.429|480.664| 17.152| -31.593| 8| 19.588
15|O16Si8 | 24|423.469|480.664| 18.892| -31.639| 8| 15.179
16|O16Si8 | 24|450.716|480.664| 17.750| -31.628| 8| 16.219
17|O16Si8 | 24|331.528|480.664| 24.131| -31.642| 8| 14.857
18|O16Si8 | 24|458.573|480.664| 17.445| -31.635| 8| 15.572
19|O16Si8 | 24|359.298|480.664| 22.266| -31.655| 8| 13.636
20|O16Si8 | 24|464.264|480.664| 17.232| -31.612| 8| 17.750
Rows: 331171 (showing first 20)
Keys: density, energy_per_tsite, n_tsites, relative_energy

ase db IZA_NNPscan.db -c id,formula,natoms,volume,mass,density,energy_per_tsite,n_tsites,relative_energy,iza_code
# Output
id|formula |natoms| volume| mass|density|energy_per_tsite|n_tsites|relative_energy|iza_code
1|O16Si8 | 24| 435.488| 480.664| 18.370| -31.676| 8| 11.594|ABW
2|O32Si16 | 48| 961.419| 961.328| 16.642| -31.645| 16| 14.612|ACO
3|O96Si48 | 144|3154.579|2883.984| 15.216| -31.664| 48| 12.810|AEI
4|O80Si40 | 120|2102.921|2403.320| 19.021| -31.703| 40| 9.021|AEL
5|O96Si48 | 144|2417.286|2883.984| 19.857| -31.666| 48| 12.586|AEN
6|O144Si72| 216|4075.300|4325.976| 17.667| -31.674| 72| 11.831|AET
7|O96Si48 | 144|2786.810|2883.984| 17.224| -31.675| 48| 11.716|AFG
8|O48Si24 | 72|1400.247|1441.992| 17.140| -31.690| 24| 10.268|AFI
9|O64Si32 | 96|1764.823|1922.656| 18.132| -31.653| 32| 13.809|AFN
10|O80Si40 | 120|2080.330|2403.320| 19.228| -31.707| 40| 8.632|AFO
11|O64Si32 | 96|2097.384|1922.656| 15.257| -31.655| 32| 13.622|AFR
12|O112Si56| 168|3820.116|3364.648| 14.659| -31.650| 56| 14.150|AFS
13|O144Si72| 216|4732.720|4325.976| 15.213| -31.664| 72| 12.793|AFT
14|O60Si30 | 90|1897.074|1802.490| 15.814| -31.659| 30| 13.268|AFV
15|O96Si48 | 144|3154.885|2883.984| 15.214| -31.664| 48| 12.776|AFX
16|O32Si16 | 48|1137.335| 961.328| 14.068| -31.591| 16| 19.790|AFY
17|O48Si24 | 72|1283.812|1441.992| 18.694| -31.620| 24| 17.034|AHT
18|O96Si48 | 144|2479.287|2883.984| 19.360| -31.681| 48| 11.155|ANA
19|O64Si32 | 96|1797.086|1922.656| 17.807| -31.662| 32| 12.924|APC
20|O64Si32 | 96|1751.393|1922.656| 18.271| -31.678| 32| 11.422|APD
Rows: 236 (showing first 20)
Keys: density, energy_per_tsite, iza_code, n_tsites, relative_energy

# Filtering of the database, e.g., for structures with relative energies < 10 kJ/(mol Si)
ase db IZA_NNPscan.db relative_energy\<10 -c density,energy_per_tsite,n_tsites,relative_energy,iza_code
# Output
density|energy_per_tsite|n_tsites|relative_energy|iza_code
19.021| -31.703| 40| 9.021|AEL
19.228| -31.707| 40| 8.632|AFO
19.385| -31.695| 24| 9.802|ATV
18.778| -31.702| 34| 9.061|DOH
19.570| -31.693| 24| 9.959|EWO
18.401| -31.698| 32| 9.451|GON
18.551| -31.695| 112| 9.807|IHW
17.778| -31.693| 288| 9.972|IMF
19.154| -31.695| 6| 9.762|JBW
18.187| -31.695| 96| 9.734|MFI
19.278| -31.709| 48| 8.443|MRE
18.035| -31.698| 90| 9.481|MSO
20.417| -31.724| 44| 7.003|MTF
19.227| -31.704| 136| 8.898|MTN
18.542| -31.693| 28| 9.966|MTW
19.137| -31.695| 60| 9.798|PCR
20.037| -31.709| 144| 8.464|PSI
18.843| -31.703| 64| 9.004|SAF
18.371| -31.703| 112| 8.975|STO
19.894| -31.706| 17| 8.671|VET
Rows: 20 (showing first 20)
Keys: density, energy_per_tsite, iza_code, n_tsites, relative_energy

The quantities shown above are available with the keys (besides standard ASE database keys):

Key
Quantity
Unit

id
Identifier

formula
Chemical formula of the unit cell

natoms
Number of atoms

volume
Unti cell volume
Å3

mass
Atomic mass of the unit cell
amu

density
Framework density
Si/nm3

energy_per_tsite
NNPscan energy
eV

n_tsites
Number of T-sites

relative_energy
Energy with respect to quartz
kJ/(mol Si)

iza_code
only for 'IZA_NNPscan.db'

 Comma separated csv files for the quantities listed above: "DEEM_NNPscan.csv" and  "IZA_NNPscan.csv"

2. Neural network potentials (NNP) for silica

SchNet [Schütt18,Schütt19] NNP files trained on DFT data at the PBE+D3 (NNPpbe) and SCAN+D3 level (NNPscan)

Simulations can be performed using SchNetPack with its ASE calculator

This example shows a simple single-point calculation

import ase.io
import torch
from schnetpack.interfaces import SpkCalculator
from schnetpack.environment import AseEnvironmentProvider

# check if GPU(s) are available
if torch.cuda.is_available():
device = "cuda"
else:
device = "cpu"

# load the NNP model
model = torch.load('SiOscan1', map_location=device)

# read some structure
atoms = ase.io.read( ... )

# define SchNetPack calculator
calc = SpkCalculator(model=model,
device=device,
energy='energy',
forces='forces',
environment_provider=AseEnvironmentProvider(6.)
)

# attach calculator to atoms object
atoms.set_calculator(calc)

# perform simulations, e.g., single-point calculation
energy = atoms.get_potential_energy()
print(energy)

3. Test set used for accuracy evaluation (ASE database: test_set_NNPscan.db)

## 数据集内容

1. 沸石数据库

Deem数据库包含331170个假设性沸石骨架，已在NNPscan级别下完成几何结构优化[Deem09, Pophale11]，需注意该数据库首行数据为α-石英（alpha-quartz），对应数据文件为`DEEM_NNPscan.db`。

另有收录国际沸石协会（International Zeolite Association, IZA）公布的236种已合成沸石骨架的数据库，同样在NNPscan级别下完成结构优化，对应数据文件为`IZA_NNPscan.db`。

上述两个数据库均为原子模拟环境（Atomic Simulation Environment, ASE）的SQLite格式数据库文件，内置带有NNPscan级别能量与力数据的ASE Atoms对象，可通过ASE的输入输出（I/O）模块读取。

此外，可通过以下查询指令提取相关参数（更多使用说明请执行`ase db --help`查看）：

bash
ase db DEEM_NNPscan.db -c id,formula,natoms,volume,mass,density,energy_per_tsite,n_tsites,relative_energy

输出示例：

标识符|晶胞化学式|原子数| 晶胞体积| 总原子质量|骨架密度|单T位点NNPscan能量|T位点数|相对能量
1|O6Si3 | 9|111.161|180.249| 26.988| -31.796| 3| 0.000
2|O16Si8 | 24|433.858|480.664| 18.439| -31.638| 8| 15.265
...（共331171条数据，此处展示前20条）

可用键值：密度、单T位点NNPscan能量、T位点数、相对能量。

bash
ase db IZA_NNPscan.db -c id,formula,natoms,volume,mass,density,energy_per_tsite,n_tsites,relative_energy,iza_code

输出示例：

标识符|晶胞化学式|原子数| 晶胞体积| 总原子质量|骨架密度|单T位点NNPscan能量|T位点数|相对能量|IZA代码
1|O16Si8 | 24| 435.488| 480.664| 18.370| -31.676| 8| 11.594|ABW
2|O32Si16 | 48| 961.419| 961.328| 16.642| -31.645| 16| 14.612|ACO
...（共236条数据，此处展示前20条）

可用键值：密度、单T位点NNPscan能量、IZA代码、T位点数、相对能量。

### 数据库筛选示例
以筛选相对能量小于10 kJ/(mol Si)的结构为例：
bash
ase db IZA_NNPscan.db relative_energy<10 -c density,energy_per_tsite,n_tsites,relative_energy,iza_code

输出示例：

骨架密度|单T位点NNPscan能量|T位点数|相对能量|IZA代码
19.021| -31.703| 40| 9.021|AEL
19.228| -31.707| 40| 8.632|AFO
...（共20条数据）

### 可用参数详情表
| 键名 | 说明 | 单位 |
|----------------------|--------------------------|----------------|
| id | 数据标识符 | — |
| formula | 晶胞化学式 | — |
| natoms | 晶胞内总原子数 | — |
| volume | 晶胞体积 | ų |
| mass | 晶胞总原子质量 | amu |
| density | 骨架密度 | Si/nm³ |
| energy_per_tsite | 单T位点NNPscan能量 | eV |
| n_tsites | T位点数 | — |
| relative_energy | 相对于α-石英的能量 | kJ/(mol Si) |
| iza_code | 仅`IZA_NNPscan.db`包含 | — |

此外还提供了上述参数的逗号分隔CSV格式文件：`DEEM_NNPscan.csv`与`IZA_NNPscan.csv`。

2. 二氧化硅神经网络势（Neural Network Potential, NNP）

包含基于PBE+D3（对应NNPpbe）与SCAN+D3（对应NNPscan）级别密度泛函理论（Density Functional Theory, DFT）数据训练得到的SchNet[Schütt18, Schütt19]神经网络势文件。可通过SchNetPack及其集成的ASE计算器开展分子模拟计算。以下为简单单点能计算示例代码：

python
import ase.io
import torch
from schnetpack.interfaces import SpkCalculator
from schnetpack.environment import AseEnvironmentProvider

# 检测GPU可用性
if torch.cuda.is_available():
device = "cuda"
else:
device = "cpu"

# 加载神经网络势模型
model = torch.load('SiOscan1', map_location=device)

# 读取目标结构
atoms = ase.io.read( ... )

# 定义SchNetPack计算器
calc = SpkCalculator(model=model,
device=device,
energy="energy",
forces="forces",
environment_provider=AseEnvironmentProvider(6.)
)

# 将计算器绑定至Atoms对象
atoms.set_calculator(calc)

# 执行模拟计算，例如单点能计算：获取势能并打印
energy = atoms.get_potential_energy()
print(energy)

3. 精度评估测试集（ASE数据库：test_set_NNPscan.db）