Browser implementation of GFN2-xTB and CREST

Abstract

Mol2Mat brings GFN2-xTB, ASE geometry optimization and CREST conformer search into the browser, executing 3D preparation, electronic-property calculation, geometry optimization and conformer sampling on the user’s device. This work evaluates numerical agreement, parallel performance and computational scale against native implementations on the same machine.

All 150 browser GFN2 results pass independent native evaluation at identical coordinates, with a maximum energy difference of 1.124e-07 Ha. Every geometry optimization matches native convergence status and step count. Complete CREST quick searches retain identical ensemble files, call counts and seed sequences across browser concurrency settings, reaching a maximum speedup of 5.44×.

The experiment comprises 354 new executions: GFN2 tasks on molecules with 9–122 atoms, complete CREST searches for three molecules, and parallel settings of 1, 2, 4, 8 and 10. The results support retention of the original scientific methods in the browser and reproducible performance gains through trajectory-level parallelism.

Numerical agreement

Native tblite independently evaluates every browser single-point and optimization result at its full-precision final coordinates. Maximum differences are 1.124e-07 Ha in energy (7.050e-05 kcal/mol), 3.736e-06 Ha/bohr in gradient components, 1.823e-09 e in charges and 7.631e-09 D in dipole vectors. All four meet the established port-consistency criteria.

Original ASE BFGS optimization from identical starting coordinates produces matching convergence status and step counts in every native/browser pair. The maximum coordinate difference across recorded optimization frames is 7.500e-05 Å. Agreement in energy, forces and optimization paths demonstrates retention of native numerical behavior for these molecules. Frames are recorded to six decimal places; final-property verification uses full-precision coordinates.

CREST parallel performance

Browser CREST distributes complete MTD/MD trajectories and optimization batches while retaining the iMTD-GC quick workflow: trial and production metadynamics, molecular dynamics, optimization, genetic crossing and CREGEN processing. Trajectories use independent electronic states and fixed random streams, with results joined in logical order at stage boundaries.

At concurrency 8, ethanol falls from 27.52 s to 6.77 s (4.07×), and n-butane from 70.18 s to 14.15 s (4.96×). At concurrency 10, n-hexane falls from 244.88 s to 44.98 s (5.44×). Times are medians of three repetitions and include complete search and result processing.

All 15 browser searches for each molecule produce byte-identical complete ensemble XYZ files, matching call counts and actual seed sequences. The 45 searches execute 690,735 search-stage GFN2 energy/gradient calls. Parallel gains result from concurrent execution of the same scientific work. The native control uses CREST 3.0.2, a matched random stream and one thread, retaining its SCC cache chain; every browser setting uses the same independent-trajectory protocol. Native timings provide implementation context; speedups compare browser concurrency settings.

Molecular geometry calculations

Aspirin, caffeine and n-tetracontane represent a multifunctional aromatic molecule, a nitrogen heterocycle and a long-chain system. Browser execution of original ASE BFGS optimization converges in 56, 12 and 18 steps for these 21-, 24- and 122-atom structures. The models use measured final coordinates and retain energy, maximum force and optimization step counts.

Molecular preparation, GFN2 optimization, electronic-property analysis and CREST search share a molecular object. Measured structures can be rotated, exported as XYZ and passed to subsequent calculations through “Use in…”. Browser computation therefore connects search, structure inspection and conformer analysis without configuring a local Python or Fortran runtime.

Ethanol: 9 atoms, 11 BFGS steps; -11.39431527 Ha. Measured XYZ

Aspirin: 21 atoms, 56 BFGS steps; -39.63155591 Ha. Measured XYZ

Caffeine: 24 atoms, 12 BFGS steps; -42.15385498 Ha. Measured XYZ

n-Eicosane: 62 atoms, 6 BFGS steps; -64.27996323 Ha. Measured XYZ

n-Tetracontane: 122 atoms, 18 BFGS steps; -127.54814429 Ha. Measured XYZ

Conformer search results

Browser searches yield 3, 2 and 26 CREGEN candidates for ethanol, n-butane and n-hexane. Minimum electronic energies differ from the native searches by 1.00e−08, 2.80e−07 and 2.70e−07 Ha. Each spectrum is referenced to its runtime’s minimum and displays the complete candidate set.

All 31 browser candidates undergo independent native tblite evaluation at identical coordinates, with a maximum energy difference of 5.137e-09 Ha and maximum forces no greater than 0.01781 eV/Å. Dynamic sampling is sensitive to electronic-state continuation and small numerical differences, producing different candidate sets across runtimes; the complete set remains identical across browser concurrency settings.

GFN2-xTB performance

Median browser single-Worker geometry-optimization compute times are 0.090 s for ethanol, 1.226 s for aspirin, 0.414 s for caffeine, 0.704 s for n-eicosane and 6.038 s for n-tetracontane. Original GFN2 and ASE computation is therefore in the subsecond-to-second range for the tested small molecules.

At child-pool concurrency 4, n-tetracontane optimization takes 5.475 s, improving on serial execution by 9.3%. Orbital-pair integrals are the parallel portion of a GFN2 calculation; ASE control, SCC and matrix operations retain serial costs. Small tasks favor serial execution, while larger structures can benefit from moderate concurrency. Native single-thread execution is faster for these tasks; the browser implementation additionally provides installation-free, on-device calculation within a continuous molecular workflow.

Parallel configurations

Complete curves for settings 1, 2, 4, 8 and 10 show that optimal concurrency depends on task structure. n-Tetracontane optimization takes 5.626 and 5.563 s at 8 and 10, with 4 fastest on this machine. CREST’s multiple trajectories gain most at 8–10. Task-level parallelism offers greater scaling potential than parallelism within an individual electronic calculation.

Resource use

Independent trajectory Workers each hold scientific runtime state, enabling multicore execution with an increased memory requirement. Chrome process-tree RSS for CREST at 8–10 is approximately 4.7–5.7 GiB, making available device memory relevant to concurrency selection. Browser values sum process-tree RSS sampled every 250 ms; native values are peak RSS for one process. Their accounting boundaries are retained separately.

Timing starts after engine assets are preloaded, with a fresh parent Worker and child pool for each run. GFN2 compute time includes ASE, electronic properties and initial child-pool startup; whole-task time adds RDKit and parent-module preparation. Initialization is recorded separately, with 169.3 MiB of deduplicated preloaded assets. CSV records also retain CPU seconds, memory and host load.

Methods

All runs use the same Apple M1 Pro (8 performance cores, 2 efficiency cores, 32 GiB RAM) and identical explicit-hydrogen coordinates prepared by browser RDKit ETKDGv3/MMFF94. GFN2 uses tblite 0.7.0 and original ASE 3.29.0, with gas-phase neutral closed-shell inputs, double precision, accuracy 1.0, 300 K electronic temperature and a 0.05 eV/Å BFGS threshold. Every optimization converges before the 1000-step cap.

Each configuration has three repetitions; figures show medians and minimum-to-maximum ranges. Settings 1/2/4 and the additional 8/10 are run in separate batches with rotating within-batch order and sequential scientific jobs. Background services remain active and host load is recorded per run. Native GFN2 sets OpenMP/BLAS thread environments; browser setting 1 is the serial parent path, while 2/4/8/10 count orbital-pair child-pool concurrency.

Native CREST is built from matched source with RNG entry points using the browser’s musl LCG, seed resets and buffered sampling. Native retains SCC state chaining; the browser uses an independent-trajectory state protocol. Complete quick stages, durations, precision and convergence conditions are retained. The reproduction package includes fixed coordinates, individual records, random streams, complete logs, native source and browser asset hashes.

Data and code

Measured records, independent calculations, input structures, experiment scripts and vector figures are available for inspection and reproduction.

Experimental conditions

Hardware and OS

Apple M1 Pro · 10 CPU cores · 32 GiB RAM; macOS 27.0 (26A428)

Browser

Chrome 154.0.8037.98, real headless browser and Workers; JSPI available.

GFN2 scientific software

tblite 0.7.0 (native OpenBLAS 0.3.33); ASE 3.29.0; NumPy 2.4.6 (native Accelerate); browser RDKit 2026.03.6.

Native CREST

CREST 3.0.2 matched-source control, GNU Fortran 16.2.0 and OpenBLAS; --gfn2 --quick --T 1 --chrg 0 --uhf 0.

CREST conditions

Accuracy 1.0; electronic temperature 300 K; 500 SCF iterations; original quick duration and all stages retained.

Random control

RDKit seed 20261003; CREST musl LCG, reset seeds 20261003 + 37×i. Native SCC state chaining differs from the browser independent-trajectory protocol.

Repetition and scale

Three runs per configuration; GFN2 9/21/24/62/122 atoms; CREST 9/14/20 atoms.

Resource sampling

Chrome process-tree RSS every 250 ms; native peak process RSS; CPU seconds and host load retained, with different accounting boundaries.

摘要

Mol2Mat 将 GFN2-xTB、ASE 几何优化与 CREST 构象搜索移植到浏览器，在用户设备上执行从三维结构准备到电子性质计算、几何优化和构象采样的完整流程。本工作通过同机原生对照，评估这一实现的数值一致性、并行性能与计算规模。

150 个浏览器 GFN2 结果全部通过原生同坐标复算，最大能量差为 1.124e-07 Ha，全部几何优化的收敛状态与步数一致。完整 CREST quick 搜索在不同并发配置下保持构象文件、调用数与种子序列一致，最高获得 5.44 倍加速。

实验共新运行 354 次，覆盖 9–122 原子的 GFN2 任务、三个分子的完整 CREST 搜索，以及 1、2、4、8、10 的并行配置。结果支持在浏览器中保留原有科学方法，并通过轨迹级并行获得可复现的性能收益。

数值一致性

每个浏览器单点与优化结果均在完整双精度最终坐标上由原生 tblite 独立复算。最大能量差为 1.124e-07 Ha，相当于 7.050e-05 kcal/mol；最大梯度分量差为 3.736e-06 Ha/bohr，电荷与偶极差分别为 1.823e-09 e 和 7.631e-09 D。四项指标均满足既有的移植一致性标准。

从相同起始坐标执行原版 ASE BFGS，两端所有优化的收敛状态和步数相同，逐步轨迹的最大坐标差为 7.500e-05 Å。能量、力和优化路径的一致性表明，浏览器实现保留了本组分子的原生数值行为。轨迹按六位小数记录，最终电子性质复算使用完整精度坐标。

CREST 并行性能

浏览器 CREST 在完整 MTD／MD 轨迹和优化批次之间分配工作，保留 iMTD-GC quick 的试探与正式元动力学、分子动力学、优化、遗传交叉及 CREGEN 整理。各轨迹使用独立电子状态与固定随机流，阶段结束后按逻辑编号汇合结果。

乙醇在 8 路并发下由 27.52 s 降至 6.77 s（4.07×）；正丁烷在 8 路下由 70.18 s 降至 14.15 s（4.96×）；正己烷在 10 路下由 244.88 s 降至 44.98 s（5.44×）。耗时为三次重复的中位数，包含完整搜索与结果处理。

每个分子的 15 次浏览器搜索均产生逐字节一致的完整构象 XYZ，并保留相同调用数与实际种子序列。45 次搜索累计执行 690,735 次搜索阶段 GFN2 能量／梯度调用。并行收益来自相同计算工作的并发执行。原生对照使用 CREST 3.0.2、匹配随机流与 1 线程，保留原 SCC 缓存链；浏览器各并发使用同一独立轨迹协议。图中原生耗时供实现对照，加速比由浏览器各并发之间计算。

分子结构计算

阿司匹林、咖啡因和正四十烷分别代表多官能团芳香分子、含氮杂环和长链体系。浏览器完成 21、24、122 原子的原版 ASE BFGS 优化，分别在 56、12、18 步收敛。下方模型直接读取本次计算的最终坐标，旁列能量、最大力与优化步数。

结构准备、GFN2 优化、电子性质分析与 CREST 搜索共用分子对象。实测结构可旋转查看、导出 XYZ，并通过「使用于…」进入后续计算；浏览器计算因此可以直接衔接检索、结构检查和构象分析，无需配置本地 Python 或 Fortran 运行环境。

乙醇: 9 atoms, 11 BFGS steps; -11.39431527 Ha. Measured XYZ

阿司匹林: 21 atoms, 56 BFGS steps; -39.63155591 Ha. Measured XYZ

咖啡因: 24 atoms, 12 BFGS steps; -42.15385498 Ha. Measured XYZ

正二十烷: 62 atoms, 6 BFGS steps; -64.27996323 Ha. Measured XYZ

正四十烷: 122 atoms, 18 BFGS steps; -127.54814429 Ha. Measured XYZ

构象搜索结果

乙醇、正丁烷与正己烷的浏览器搜索分别输出 3、2、26 个 CREGEN 候选。与原生搜索的最低电子能量差分别为 1.00e−08、2.80e−07、2.70e−07 Ha。下图按各端最低能量归零，显示完整候选集合的相对电子能量。

全部 31 个浏览器候选均由原生 tblite 在相同坐标上独立复算，最大能量差为 5.137e-09 Ha，候选最大力不超过 0.01781 eV/Å。动力学采样对电子状态延续和微小数值差异敏感，两端的候选集合有所不同；各浏览器并发配置的完整集合保持一致。

GFN2-xTB 性能

浏览器单 Worker 几何优化的计算中位数为：乙醇 0.090 s、阿司匹林 1.226 s、咖啡因 0.414 s、正二十烷 0.704 s、正四十烷 6.038 s。对受测的常见小分子，原版 GFN2 与 ASE 计算已进入亚秒至秒级。

正四十烷优化在 4 路子池下为 5.475 s，较串行改善 9.3%。单次 GFN2 的并行部分是轨道对积分，ASE 控制、SCC 与矩阵运算仍有串行成本；小任务更适合串行执行，较大结构可受益于适度并发。原生单线程在这组任务中耗时更短，浏览器实现同时提供免安装、设备内计算与连续的分子工作流。

并行配置

1、2、4、8、10 配置的完整曲线显示，最佳并发取决于任务结构。正四十烷几何优化的 8、10 路中位数分别为 5.626、5.563 s，本机以 4 路最快；CREST 多轨迹搜索则在 8–10 路获得最大收益。任务级并行比单次电子计算内部并行具有更大的扩展空间。

资源开销

独立轨迹 Worker 各自持有科学运行状态，支持多核执行并增加内存需求。8–10 路 CREST 的 Chrome 进程树 RSS 约为 4.7–5.7 GiB，适合结合设备内存选择计算并发。浏览器数值为 250 ms 采样的进程树 RSS 总和，原生数值为单个进程峰值；两者统计范围分别列示。

计时在引擎资源预载后开始，每次新建主 Worker 与子池。GFN2 计算区间包含 ASE、电子性质与首次子池启动，完整任务另含 RDKit 和主计算模块准备。初始化单独记录，去重预载资源为 169.3 MiB；CSV 同时提供 CPU 秒、内存和主机负载。

实验方法

实验在同一台 Apple M1 Pro（8 性能核、2 能效核，32 GiB 内存）上运行。两端使用浏览器 RDKit ETKDGv3／MMFF94 生成的同一组显式含氢坐标。GFN2 对照为 tblite 0.7.0 与原版 ASE 3.29.0；统一气相中性闭壳层、双精度、accuracy 1.0、300 K 电子温度及 0.05 eV/Å BFGS 收敛阈值。全部优化在 1000 步上限前收敛。

各配置重复三次，图中为中位数与最小至最大值。1／2／4 和追加的 8／10 配置分批运行，批内顺序跨轮轮换，科学任务顺序执行。系统后台服务保持运行，主机负载逐次记录。原生 GFN2 设置 OpenMP／BLAS 线程环境；浏览器 1 为主 Worker 串行路径，2／4／8／10 为轨道对子池并发。

CREST 原生对照按匹配源码构建，随机数入口使用与浏览器一致的 musl LCG、种子重置与缓冲抽样。原生沿用 SCC 状态链，浏览器采用独立轨迹状态协议；完整 quick 的阶段、时长、精度与收敛条件均保留。复现包保存固定坐标、逐次记录、随机流、完整日志、原生源码和浏览器资产哈希。

数据与代码

实测记录、独立复算、输入结构、实验脚本与矢量图表可下载，用于检查结果和复现实验。

实验条件

硬件与系统

Apple M1 Pro · 10 CPU cores · 32 GiB RAM；macOS 27.0 (26A428)

浏览器

Chrome 154.0.8037.98，真实 headless 浏览器与 Worker，JSPI 可用。

GFN2 科学软件

tblite 0.7.0（原生 OpenBLAS 0.3.33）；ASE 3.29.0；NumPy 2.4.6（原生 Accelerate）；浏览器 RDKit 2026.03.6。

原生 CREST

CREST 3.0.2 匹配源码控制，GNU Fortran 16.2.0、OpenBLAS；--gfn2 --quick --T 1 --chrg 0 --uhf 0。

CREST 条件

accuracy 1.0；电子温度 300 K；SCF 上限 500；原 quick 时长与全部阶段保留。

随机控制

RDKit seed 20261003；CREST musl LCG，重置种子 20261003 + 37×i。原生 SCC 状态链与浏览器独立轨迹协议不同。

重复与规模

每项配置 3 次；GFN2 9／21／24／62／122 原子；CREST 9／14／20 原子。

采样边界

Chrome 进程树 RSS 每 250 ms；原生进程峰值 RSS；CPU 秒与主机负载保留，资源边界不同。