[论文解读] On data and dimension in chemistry -- irreversibility, concealment and emergent conservation laws

Alex Blokhuis, Martijn van Kuppeveld+2 more|arXiv (Cornell University)|Jun 15, 2023

Advanced Thermodynamics and Statistical Mechanics被引用 5

一句话总结

本文提出共生产律 $\Upsilon = \ell_{\bullet} + \wedge_{\bullet}$，用以解释由于共线且不可逆反应导致的化学反应网络（CRNs）中涌现的非整数守恒律。该理论解决了长期以来关于机器发现的表观守恒律的疑难问题，表明该守恒律源于共生产而非内在守恒，并推导出隐藏物种与不可区分物种的维数定律，使得仅凭光谱数据即可推断CRN的结构。

ABSTRACT

Chemical systems are interpreted through the species they contain and the reactions they may undergo, i.e., their chemical reaction network (CRN). In spite of their central importance to chemistry, the structure of CRNs continues to be challenging to deduce from data. Although there exist structural laws relating species, reactions, conserved quantities and cycles, there has been limited attention to their measurable consequences. One such is the dimension of the chemical data: the number of independent reactions or equivalently independent species, which corresponds to the number of measured variables minus the number of constraints. In this paper we attempt to relate the experimentally observed dimensional features to conservation laws and underlying CRN structure. Our approach extends to any Markov model as well as many nonlinear models in statistical physics and furnishes new analytical tools to find exact solutions. In particular, we investigate the effects of species that are concealed and reactions that are irreversible. For instance, irreversible reactions can have proportional rates. The resulting reduction in degrees of freedom can be captured by the co-production law relating co-production relationships to emergent non-integer conservation laws and broken cycles. This law resolves a recent conundrum posed by a machine-discovered candidate for a non-integer conservation law, and characterizes certain types of CRN behavior. We also obtain laws that allow us to relate data dimension to network structure in cases where some species cannot be discerned or distinguished by a given analytical technique, allowing to narrow down candidate CRNs from experimental data more effectively.

研究动机与目标

• 解决近期在机器学习CRN模型中发现的异常非整数守恒律的起源问题。
• 形式化实验观测数据维度与底层CRN结构之间的关系，尤其在不可逆性与隐藏性条件下的关系。
• 推导出包含隐藏物种与不可区分物种系统的维数定律，使能从不完整或间接测量中推断CRN拓扑结构。
• 阐明光谱特征（如等吸光点）与局部守恒律及网络拓扑之间的关系。
• 提供一个理论框架，简化CRN重构过程，类比于分子光谱学中的逻辑谜题。

提出的方法

• 推导共生产律 $\Upsilon = \ell_{\bullet} + \wedge_{\bullet}$，其中 $\Upsilon$ 为共生产指数，$\ell_{\bullet}$ 表示涌现的非整数守恒律，$\wedge_{\bullet}$ 衡量由于共线性导致的循环破坏程度。
• 应用化学计量矩阵分析建模不可逆反应，识别共线反应对（如 $r_4$ 与 $r_5$），这些反应会减少系统的自由度。
• 利用连续培养模型与物种隐藏模型推导维数定律，表明可观测物种的损失会降低有效数据维度。
• 引入化学计量矩阵的变换方法，将共线反应合并为一个具有速率依赖化学计量的等效反应。
• 通过计算与数值观测到的守恒律 $CQ_3$ 误差在1%以内的左零空间向量 $\pmb{\ell}_{\bullet}$，验证了共生产律的有效性。
• 将该框架应用于真实大气CRN模型，确认 $CQ_3$ 并非真正的守恒律，而是共生产的涌现结果。

Figure 1: a-b) CRNs whose conservation laws follow from integer stoichiometry. c) conservation laws. Making reactions $r_{1},r_{2}$ irreversible, an additional non-integer conservation conserved quantity $L^{(3)}$ emerges. d) CRN with an emergent non-integer conserved quantity. For integer stoichiom

实验结果

研究问题

• RQ1 在CRNs中，非整数守恒律的出现原因是什么，尤其是在通过机器学习算法检测到的情况下？
• RQ2 具有成比例速率的不可逆反应如何影响化学反应网络的维度与结构推断？
• RQ3 当部分物种被隐藏或不可区分时，能在多大程度上从不完整或间接的光谱数据中推断CRN的结构？
• RQ4 光谱数据中的等吸光点如何与局部守恒律及网络拓扑相关联？
• RQ5 共生产律 $\Upsilon = \ell_{\bullet} + \wedge_{\bullet}$ 是否能够解释并解决机器发现的守恒律中的异常现象？

主要发现

• 共生产律 $\Upsilon = \ell_{\bullet} + \wedge_{\bullet}$ 有效解释了异常守恒律 $CQ_3$ 的起源，表明其源于共线不可逆反应导致的非整数守恒，而非内在守恒。
• 左零空间向量 $\pmb{\ell}_{\bullet} = (6,-5,1,3,9,6,3,6,4,-3,2.18)$ 的数值与观测到的 $CQ_3 \approx (6,-5,1,3,9,6,3,6,4,-3,2.21)$ 非常接近，确认了该定律的准确性。
• 在大气CRN模型中，将共线反应 $r_4$ 与 $r_5$ 合并为一个具有速率依赖化学计量（$p = 0.40541$）的等效反应，降低了系统的自由度，并促成 $\ell_{\bullet} = 1$ 的出现。
• 本研究证实 $CQ_3$ 并非真正的守恒律，而是共生产的涌现现象，从而解决了近期在机器学习CRN分析中出现的疑难问题。
• 该框架通过将等吸光点与光谱不变性关联到局部守恒律及网络拓扑，实现了从光谱数据重构CRN结构。
• 针对隐藏物种与不可区分物种的维数定律，使得即使实验数据未能完全分辨所有物种，也能推断CRN结构，显著简化了网络重构过程。

Figure 2: (a) Simulated absorption spectra for CRN $\ce{C}\leftrightarrows\ce{A}+\ce{B}\leftrightarrows\ce{D}$ , no isosbestic point ( $d(\Lambda)=2$ ). b) simulated spectra for CRN $\ce{C}\leftarrow\ce{A}+\ce{B}\rightarrow\ce{D}$ , a close-up reveals an isosbestic point ( $d(\Lambda)=1$ ). (c) Unde