跳转至

域门控潜在扩散:基于第一性原理验证的 HMX 级含能材料生成式逆向设计

文章背景与核心概要

含能材料作为推进、爆破和安全系统的核心,其下一代化合物的设计需要在“高能量释放”、“对意外引爆的低敏感性”以及“可行的化学合成路线”这三个相互冲突的需求之间取得平衡。探索这一庞大的分子空间通常依赖于生成式人工智能模型,但标准模型面临着现有化学数据严重不可靠的挑战:在大约 6,600 个记录在案的分子中,仅有约 3,000 个通过第一性原理进行了测量或计算。

为了克服这一痛点,作者引入了域门控潜在扩散(Domain-Gated Latent Diffusion, DGLD)模型。DGLD 将数据的可靠性作为核心设计参数,将标签细分为四个可信度层级。高可信度数据直接引导生成过程,而更大规模、可靠性较低的数据则用于训练模型的结构合理性。每个生成的分子均需通过严格的四阶段筛选流水线,并在最终阶段接受量子化学密度泛函理论(DFT)的审核。DGLD 成功识别出 10 个 PubChem 中未收录的新分子,其中最佳候选分子——3,4,5-三硝基-1,2-异恶唑——在爆轰性能上可与 HMX 和 PETN 等基准炸药相媲美,同时具备一条切实可行的四步合成路线。


Executive Summary

Designing next-generation energetic materials—essential for propulsion, demolition, and safety systems—requires balancing three conflicting requirements: high energy release, low sensitivity to accidental initiation, and a practical chemical synthesis route. Exploring this astronomical molecular space typically relies on generative AI models. However, standard models struggle because the vast majority of existing chemical data is unreliable: out of ~66,000 recorded molecules, only ~3,000 have properties measured or computed from first principles.

To overcome this, the authors introduce Domain-Gated Latent Diffusion (DGLD). DGLD treats data reliability as a core design parameter by sorting labels into four trust tiers. Trustworthy data directly steers the generation process, while the larger, less reliable dataset trains the model on structural plausibility. Every generated molecule passes through a rigorous four-stage screening pipeline culminating in a quantum-chemical Density Functional Theory (DFT) audit. DGLD successfully identified 10 novel molecules unknown to PubChem, with the top candidate—3,4,5-trinitro-1,2-isoxazole—matching benchmark explosives like HMX and PETN in detonation performance while featuring a viable four-step synthesis route.

设计下一代含能材料(对于推进、爆破和安全系统至关重要)需要在三个相互冲突的要求之间取得平衡:高能量释放对意外引爆的低敏感性以及实用的化学合成路线。探索这一庞大的分子空间通常依赖于生成式人工智能模型。然而,标准模型举步维艰,因为绝大多数现有的化学数据并不可靠:在约 66,000 个记录在案的分子中,仅有约 3,000 个的属性是通过第一性原理测量或计算得出的。

为了克服这一问题,作者引入了域门控潜在扩散(Domain-Gated Latent Diffusion, DGLD)。DGLD 将数据可靠性视为核心设计参数,将标签分为四个信任层级。可信数据直接指导生成过程,而更大、可靠性较低的数据集则训练模型的结构合理性。每个生成的分子都经过严格的四阶段筛选管道,最终以量子化学密度泛函理论(DFT)审计告终。DGLD 成功识别出了 10 个 PubChem 未知的新分子,其中最佳候选分子——3,4,5-三硝基-1,2-异恶唑——在计算爆轰性能上与 HMX 和 PETN 等基准炸药相匹配,同时具有可行的四步合成路线。


Abstract

Energetic materials power mining, demolition, propulsion and airbags, yet today's compounds were designed decades ago. A successor must combine high energy release, low sensitivity to accidental initiation and a practical synthesis route, found within an astronomically large molecular space. Generative models are the natural search tool, but their training data are mostly untrustworthy: of approximately 66,000 molecules with recorded properties, only approximately 3,000 were measured or computed from first principles. Models trained on all of them imitate the rough estimates and propose molecules that collapse under real physics.

We introduce Domain-Gated Latent Diffusion (DGLD), a diffusion model that treats data reliability as an explicit design parameter: labels are sorted into four trust tiers, and only trustworthy ones steer generation, while the unreliable majority still teaches the model what a plausible molecule looks like. Learned controls tune performance, safety and viability independently, and every proposal passes a four-stage screen ending in a quantum-chemical DFT audit. DGLD proposes 10 molecules unknown to PubChem that survive this screen. The best, 3,4,5-trinitro-1,2-isoxazole, matches the benchmark explosives HMX and PETN in calculated detonation performance, is unlike molecules in its training set, and has a four-step synthesis route. Trust gating is chemistry-independent and can be applied wherever abundant weak data surround a reliable core.

含能材料为采矿、爆破、推进和安全气囊提供动力,但当今的化合物却是几十年前设计的. 其替代品必须在极其庞大的分子空间中结合高能量释放、对意外引爆的低敏感性以及实用的合成路线。生成模型是天然的搜索工具,但它们的训练数据大多不可信:在约 66,000 个具有记录属性的分子中,只有约 3,000 个是通过第一性原理测量或计算的。在所有这些数据上训练的模型会模仿粗略的估计,并提出在真实物理学下会崩溃的分子。

我们引入了域门控潜在扩散(DGLD),这是一种将数据可靠性视为显式设计参数的扩散模型:标签分为四个信任层级,只有可信的标签才能指导生成,而不可靠的大多数仍然教导模型什么是有合理性的分子。学习到的控制项可以独立调整性能、安全性和可行性,并且每个提案都通过了以量子化学 DFT 审计结束的四阶段筛选。DGLD 提出了 10 个 PubChem 未知的分子,它们通过了这一筛选。最好的一个是 3,4,5-三硝基-1,2-异恶唑,其计算出的爆轰性能与基准炸药 HMX 和 PETN 相匹配,不同于其训练集中的分子,并且具有四步合成路线。信任门控与化学性质无关,可以应用于任何丰富的弱数据围绕可靠核心的领域。


Article Details & Metadata

文章详情与元数据

license icon

相关资源与链接 * 文献计量工具: NASA ADSGoogle 学术Semantic Scholar * 许可证: 知识共享署名 4.0 (参见下方许可证图标)

license icon