分子與程序設計中優化演算法的性能比較

Abstract

新材料或新試劑的開發對於科技進步至關重要，而尋找更好的材料或試劑往往不是那麼容易，通常需要大量的實驗資源投入，而這類任務之所以困難是因為牽涉分子的選擇下導致潛在的組合變得比單純考慮操作條件時來的多非常多，因為牽涉到分子選擇與操作條件的最佳化，僅僅找到合適的分子並不足夠，還需要同時找出與該分子反應相匹配的實驗參數，才能最大限度地提升新材料或新試劑的效用，以找到最優的使用條件。 近年來，隨著計算機輔助技術的發展，已經有許多優秀的分子設計和操作條件優化的工具。這些工具都能有效地幫助實驗化學家做出較優的實驗選擇，並且取得了很好的結果。然而，現有的工具大多專注於單獨進行分子結構設計或操作條件優化，較少有人討論可以同時優化分子設計和操作條件的工具。 因此，本研究旨在開發一種適合分子設計與操作條件的協同優化工具，以提高新材料或試劑開發的效率，目標在於減少實驗資源與時間投入，並找出更好的分子結構和操作條件組合。在這個研究中，我採用了四種常用的智能演算法：基因算法（GA）、粒子群算法（PSO）、模擬退火算法（SA）以及人工蜂群算法（ABC）。這些演算法被修改並優化，以適應分子生成和實驗參數最佳化的任務需求。同時，我結合了分子生成變分自編碼器（Variational Autoencoder）和機器學習模型，對這些演算法的表現進行了比較。 結果顯示，在少量實驗數據作為起始資料的情況下，代理模型選擇基於樹的模型（tree-based model）較為合適，而人工蜂群演算法（ABC）和模擬退火算法（SA）在同時優化分子結構和操作條件方面表現尤為出色，在探索（Exploration）與利用（Exploitation）之間達到了良好的平衡，這表明即使在實驗數據點稀少的情況下，這些優化方法仍能為實驗推薦提供有力支持。

The development of new materials or reagents is crucial for technological advancement. However, finding better materials or reagents is not easy and typically requires a significant investment of experimental resources. This kind of task is challenging because the selection of molecules involved leads to a vast number of potential combinations, far exceeding the number when only considering reaction conditions. Since molecular selection and reaction condition optimization are intertwined, simply finding suitable molecules is insufficient. It is also necessary to identify experimental parameters simultaneously that match the reaction of the molecule to maximize the utility of the new material or reagent and find the optimal usage conditions. In recent years, with the development of computer-aided technologies, there have been many excellent tools for molecular design and reaction condition optimization. These tools can effectively assist experimental chemists in selecting the next round of experiments and have achieved remarkable results. However, most existing tools focus only on either optimization of molecular structure or operational conditions, with fewer studies exploring tools that can simultaneously optimize both. This study aims to develop a coordinated optimization tool tailored for molecular design and operational conditions, with the goal of improving the efficiency of new material or reagent development. The focus is on reducing the resources and time required for experiments while identifying better combinations of molecular structures and operational conditions. In this research, I employed four commonly used intelligent algorithms: Genetic Algorithm (GA), Particle Swarm Optimization (PSO), Simulated Annealing (SA), and Artificial Bee Colony (ABC). These algorithms were modified and optimized to suit the tasks of molecular generation and experimental condition optimization. Additionally, I integrated a Variational Autoencoder (VAE) for molecular generation and machine learning models to compare the performance of these algorithms. Our results demonstrate that, with a limited amount of initial experimental data, tree-based models are more suitable for surrogate modeling. Meanwhile, the Artificial Bee Colony (ABC) and Simulated Annealing (SA) algorithms excel in simultaneously optimizing molecular structures and operational conditions, achieving a good balance between exploration and exploitation. This demonstrates that even in situations with limited experimental data points, these optimization methods can still provide robust support for experimental recommendation.