人工智慧輔助綠色供應鏈風險評估： 應用於電子業供應鏈減塑策略

Abstract

全球塑膠使用量持續攀升，國際組織相繼頒布禁令與法規，促使企業對供應鏈推動減塑轉型。然而，策略施行伴隨著新的供應鏈風險，如何客觀、全面地評估上下游風險，是協助企業永續轉型的關鍵。傳統供應鏈風險評估（Supply Chain Risk Assessment, SCRA）高度仰賴專家主觀判斷，動態更新速度慢，且難以將非結構化風險資訊轉換為可量化之模型參數，導致風險涵括範圍受限，評估結果欠缺客觀性。 有鑑於此，本研究提出以大型語言模型（Large Language Models, LLM）文本挖掘與因果推理技術，輔助供應鏈風險評估，並結合供應鏈作業參考模式（Supply Chain Operations Reference, SCOR），將風險精確定位至供應鏈各作業節點。本框架並非取代傳統評估方法，而是提供一套 AI 輔助的智慧風險評估框架，探討 LLM 應用於供應鏈風險評估之可行性與邊界，以提升評估之客觀性、效率與全面性。 為驗證框架適用性，本研究以台灣電子零組件製造業電子包裝材料（Electronic Packaging Materials, EPM）減塑策略為實例，針對「重複使用塑膠」與「再生塑膠」兩項策略進行比較分析。結果顯示，重複使用塑膠策略之風險主要集中於客戶市場網絡與逆向物流回收體系之不成熟，最高曝險損失可達新台幣 414 億元，建議於回收需求辨識等節點強化資訊交換與需求透明度；再生塑膠策略之風險則集中於原料品質不穩定與供應來源混雜所引發之製程挑戰，最高曝險損失可達新台幣 722 億元，建議於物料分選節點建立材料比例監控機制。 本研究之主要學術貢獻在於整合 LLM 自然語言處理與因果推理能力至 SCOR 架構，建立從非結構化知識到量化風險模型的完整轉化路徑，填補現有文獻在 AI 輔助風險量化操作流程上的研究空白。於實務層面，本框架可協助企業精準辨識供應鏈風險熱點，依據 SCOR 最佳實務（Best Practice）制定具體之緩解策略與 KPI 追蹤指標，有效降低綠色轉型過程中的決策不確定性。

Global plastic consumption continues to rise, prompting international organizations to issue successive bans and regulations that push enterprises toward plastic-reduction transitions across their supply chains. However, the implementation of such strategies introduces new supply chain risks, and the ability to objectively and comprehensively assess upstream and downstream risks has become critical to supporting corporate sustainability transitions. Traditional Supply Chain Risk Assessment (SCRA) relies heavily on subjective expert judgment, updates slowly, and struggles to convert unstructured risk information into quantifiable model parameters, which results in limited risk coverage and a lack of objectivity in assessment outcomes. In light of this, the present study proposes leveraging Large Language Models (LLMs) for text mining and causal inference to assist supply chain risk assessment, integrated with the Supply Chain Operations Reference (SCOR) model to precisely map risks onto specific supply chain process nodes. This framework is not intended to replace traditional assessment methods, but rather to offer an AI-assisted intelligent risk assessment framework that explores the feasibility and boundaries of applying LLMs to supply chain risk assessment, thereby enhancing the objectivity, efficiency, and comprehensiveness of the evaluation process. To validate the applicability of the proposed framework, this study takes the plastic-reduction strategies for Electronic Packaging Materials (EPM) in Taiwan's electronic components manufacturing industry as a case study, conducting a comparative analysis of two strategies: reusable plastics and recycled plastics. The results indicate that the risks associated with the reusable plastics strategy are primarily concentrated in the immaturity of customer market networks and reverse logistics recycling systems, with a maximum exposure loss of up to NT$41.4 billion; it is recommended that information exchange and demand transparency be strengthened at nodes such as recycling demand identification. For the recycled plastics strategy, risks are mainly concentrated in process challenges arising from unstable raw material quality and mixed supply sources, with a maximum exposure loss of up to NT$72.2 billion; it is recommended that a material composition monitoring mechanism be established at the material sorting node. The primary academic contribution of this study lies in integrating LLM-based natural language processing and causal reasoning capabilities into the SCOR framework, establishing a complete transformation pathway from unstructured knowledge to quantified risk models, thereby filling the research gap in the existing literature regarding operational workflows for AI-assisted risk quantification. At the practical level, this framework can help enterprises precisely identify supply chain risk hotspots and, based on SCOR Best Practices, formulate concrete mitigation strategies and KPI tracking indicators, this effectively reducing decision-making uncertainty in the green transition process.