A Cross-Domain Engineer-Researcher Bridging Materials Science
and Machine Intelligence

Introduction

The boundary between computational materials science and machine learning has blurred in the past decade. Practitioners who can move fluidly between density-functional theory (DFT), Bayesian decision theory, and production-grade software engineering remain rare. This note describes such a profile [1, 2].

The subject is enrolled in the doctoral program at Kyoto University, where strain engineering of ferroelectric perovskites under active learning is the principal research thread. Concurrently, the subject is employed as an ML engineer at Athena Technologies, where industrial AI is shipped to regulated sectors.

Methods

2.1 Strain optimization via active learning

We treat the six-dimensional strain tensor as a continuous design variable and the target property (band gap, polarization, elastic modulus) as an expensive black-box function f: ℝ⁶ → ℝ.

x* = arg max_{x ∈ X} α(x; D),   D = {(xᵢ, f(xᵢ))}

The acquisition α is Expected Improvement (GP-EI) with an RBF kernel. First-principles calculations are dispatched as Slurm jobs on the KUDPC cluster [3].

2.2 Industrial AI delivery

At Athena Technologies, three engagements have been delivered: (i) a secure local LLM running entirely behind an air-gap for a major bank, supporting masking, approval-document review, and automated business document generation; (ii) anomaly detection and process optimization on factory sensor logs; (iii) a high-recall RAG assistant tuned for medical knowledge bases.

Experience

1. 銀行向けセキュアローカル LLM. インターネットから物理的に遮断した閉域ネットワーク上にローカル LLM を構築し、金融機関向け AI エージェントを開発。マスキング・稟議書レビュー・業務文書自動作成など複数のユースケースを実装。 See [4], [5].

2. 製造現場データ分析 AI. 製造ラインのセンサーデータ・業務ログを基に、異常検知や工程最適化を行う AI モデルを開発。データ分析から実装・運用まで一気通貫で担当。

3. 医療向け RAG 専属 AI アプリ. 医療現場のナレッジベースを活用した RAG による専属 AI アシスタントを開発。正確性と応答速度の両立を重視した設計を担当。

Selected Side Projects

Local Discovery. 日本中の旅行先の飲食店を共有し合う Web アプリ。投稿・発見・共有の体験をシンプルに設計。 Ruby / JavaScript / Rails [link].

Manimtube. コンピューターサイエンス解説動画を集約する学習プラットフォーム。視聴導線と検索性を改善。 Python / Next.js / PostgreSQL [link].

Starbucks Map App. スターバックス店舗を地図上で探索し、訪問スタンプを収集する Android アプリ。 Kotlin / Android / Maps API.

Pomodoro Blocker. ポモドーロ + サイトブロックを融合した Chrome 拡張。集中時間を最大化するツール。 JavaScript / Chrome Extension [link].

Awards

• Best Presentation Award. Kyoto University Joint-Symposium on Mechanics of Advanced Materials & Structures 2025 · Nov 2025.
• 京都大学 BOOST 次世代 AI 奨学金. Kyoto University BOOST — Next Generation AI Scholarship.

Discussion

Active learning bridges the gap between expensive computational materials science and tractable optimization. The same Bayesian decision machinery transfers cleanly to industrial engagements where labeled data is scarce and acquisition is expensive. Cross-domain practice keeps both halves honest.

Acknowledgments

Thanks to the Shimada Group at Kyoto University and the engineering team at Athena Technologies for the collaborations referenced throughout this profile.

References

1. Tao Xu, Shodai Nakai, Yasuaki Maruyama, Hiroki Noda, Susumu Minami, Hiroyuki Hirakata, Takahiro Shimada, “Engineering cleavage fracture of silicon via excess charge doping”, Engineering Fracture Mechanics 331, 111721, 2026.
2. 見波将, 丸山泰明, 阿部能将, 仲山智裕, 嶋田隆広, “Data-driven strain optimization of nonlinear ferroelectric properties via active ML and high-throughput DFT”, 日本機械学会論文集 91巻 941号 24-00184, 2025.
3. Maruyama, Y., Source repository, GitHub. https://github.com/maruyamayasuaki
4. Maruyama, Y., Technical writing index, Qiita. https://qiita.com/yasu_qita