Blog

Mol2Mat engineering articles on GFN2-xTB, CREST, powder XRD, VTNA, ASKCOS / Pistachio and AiZynthFinder, with numerical validation and downloadable records.

Methods, engineering and benchmarks

Browser implementation of GFN2-xTB and CREST

A same-machine benchmark of molecular geometry calculations, parallel conformer search and numerical agreement with native implementations.

GFN2-xTB 与 CREST 的浏览器实现 · 分子结构计算、并行构象搜索与原生数值一致性的同机 Benchmark。

Calculating powder X-ray diffraction in the browser

Double-precision powder diffraction with pymatgen as the reference. Shorter scheduling waits reduce the warmed silicon task median from 78.2 to 3.5 ms, with matching peak tables.

浏览器端粉末 X 射线衍射计算与参考实现对照 · 以 pymatgen 为参考实现双精度粉末衍射计算。减少调度等待后，预热硅任务的中位耗时从 78.2 ms 降至 3.5 ms，峰表保持一致。

Running VTNA kinetic analysis in the browser

Compiling the Auto-VTNA core and solvers for the browser. Curves, scores and residuals show agreement with native results and how fitting constraints affect the selected orders.

浏览器端 VTNA 动力学分析与数值一致性验证 · 将 Auto-VTNA 的数值核心和求解器编译到浏览器，以曲线、评分和残差展示原生对照，以及拟合约束对级数选择的影响。

Deploying and validating ASKCOS / Pistachio in the browser

Porting the original FP32 reaction model with step-by-step probability and search checks. Scheduling reduces two warmed task times by 28.5% and 33.8%; read-only state reuse reduces copying.

反应预测模型的浏览器部署与数值验证：ASKCOS / Pistachio · 移植原始 FP32 反应预测模型，逐步检查概率与搜索决策。调度优化使两例预热任务耗时降低 28.5% 和 33.8%，只读状态复用减少复制。

Running AiZynthFinder in the browser with sharded stock

Running the original template policy, chemistry and tree search in the browser. Stock shards provide on-demand lookup of 17.4 million full keys, with native comparisons for six target molecules.

AiZynthFinder 的浏览器移植、库存分片与原生验证 · 在浏览器运行原始模板策略、化学操作与树搜索。库存分片按需查询 1742 万项完整键，六个目标分子的结果与原生实现逐层对照。

GFN2-xTB 与 CREST 博客：工程与验证

GFN2-xTB、CREST、粉末 XRD、VTNA、ASKCOS / Pistachio 与 AiZynthFinder 工程文章，附验证数据与下载记录。