mei28 / TSUNDOKU

読んだ論文をまとめていく
1 stars 0 forks source link

prompt-engineering-of-gpt-4-for-chemical-research-what-cancannot-be-done #89

Open mei28 opened 10 months ago

mei28 commented 10 months ago

id: prompt-engineering-of-gpt-4-for-chemical-research-what-cancannot-be-done aliases:

prompt-engineering-of-gpt-4-for-chemical-research-what-cancannot-be-done

一言で言うと

LLMはどこまで化学物質を理解しているかを調査

論文リンク

著者/所属機関

投稿日付(yyyy/MM/dd)

先行研究と比べてどこがすごい?

技術・手法のキモはどこ?

どうやって有効だと検証した?


  1. LLMのどこまで知識を持っているのか?

    化合物(Compounds)

    • 一般的な科学の教科書やwebの知識がある
    • TEMPOなどの少し専門的な知識もある。これはWikiから訓練されているかも。
    • しかし、さらに派生したもの4-cyano TEMPOは持っていない。プレプリントやオープンアクセスなどはデータにはいっていない?
    • プロンプト:Tell me the phsical mand chemical properties of tolen

物理化学(Physcal chemistry)

有機化学 (Organic chemistry)


4.化学/材料情報学(Cheminformatics and materials informatics)

名前とSMILESの変換

Reasoning

特性予測 (Property prediction)

Planing (最適化-1変数)

Planing (最適化-複数変数)

Planning (ブラックボックス最適化)

Planing (分子探索)

ロボットアームの指示

LLMによる自律的な研究

コメント

次はなに読む?