"/typst/templates/paper-slides.typ": body-s = 14pt head-s = 17pt : paper-slides.with( title: "MatMind: structure-activity knowledge-driven generative foundation model for materials science", subtitle: "材料科学基座模型：构效关系 + 双头架构 + 物理强化学习", author: "Biscuit", date: "2026-07-12", ) -slide[ (center + horizon)[ (size: 28pt, weight: "bold", fill: accent)[MatMind] (0.3em) (size: 18pt, fill: accent)[Structure-Activity Knowledge-Driven Generative Foundation Model for Materials Science] (2em) (size: 13pt, fill: secondary-text)[arXiv 2606.07712 · Springer Nature 审稿中] (0.5em) (size: 13pt, fill: secondary-text)[基于S1-Base 8B（Qwen3-8B）· Alexandria + MP-20 数据库] (1.5em) ( columns: (1fr, 1fr, 1fr), gutter: 1.5em, [ -block[ (size: body-s, fill: accent, weight: "bold")[E hull MAE 0.0109] (size: 12pt, fill: secondary-text)[超越 CGCNN/M3GNet]]], [ -block[ (size: body-s, fill: accent, weight: "bold")[S.U.N. 65.3%] (size: 12pt, fill: secondary-text)[无条件生成超越 MatterGen]]], [ -block[ (size: body-s, fill: accent, weight: "bold")[4.3× 提升] (size: 12pt, fill: secondary-text)[极稀疏条件生成 (21/600k)]]], ) ] ] Outline ( columns: (1fr, 2fr), gutter: 2em, align: top, [ ( columns: (0.5fr, 4fr), inset: 6pt, stroke: none, align: (center, left), [ (size: 22pt, weight: "bold", fill: accent)[1]], [ (size: body-s)[三阶段渐进式训练 \ (fill: secondary-text, size: 12pt)[pretrain → SAR + 双头 → GRPO RL]]], [ (size: 22pt, weight: "bold", fill: accent)[2]], [ (size: body-s)[双头架构设计 \ (fill: secondary-text, size: 12pt)[语言头 + 回归头 + warm-up]]], [ (size: 22pt, weight: "bold", fill: accent)[3]], [ (size: body-s)[GRPO 多目标物理 RL \ (fill: secondary-text, size: 12pt)[组内相对优势 + 4 种奖励]]], [ (size: 22pt, weight: "bold", fill: accent)[4]], [ (size: body-s)[评价指标 \ (fill: secondary-text, size: 12pt)[MAE + S.U.N. + 评价管线]]], [ (size: 22pt, weight: "bold", fill: accent)[5]], [ (size: body-s)[实验结果 \ (fill: secondary-text, size: 12pt)[性质预测 / 无条件生成 / 条件迁移]]], [ (size: 22pt, weight: "bold", fill: accent)[6]], [ (size: body-s)[局限与思考 \ (fill: secondary-text, size: 12pt)[计算成本 / 消融待补 / 代码未开源]]], ) ], [ [ (size: body-s, weight: "bold", fill: accent)[核心问题] (0.3em) (size: body-s, fill: secondary-text)[材料 AI 被"窄架构"分割：GNN 做预测，扩散模型做生成。能否用一个 LLM 统一承载结构表示、定量预测、构效推理？] (0.5em) (size: body-s, weight: "bold", fill: accent)[MatMind 的回答] (0.2em) (size: body-s)[三阶段渐进式训练框架，协同激活领域知识与物理反馈] (0.3em) ( columns: (1fr, 1fr, 1fr), inset: 3pt, stroke: 0.4pt, align: (center, center, center), [ (size: 11pt, weight: "bold")[构效知识注入]], [ (size: 11pt, weight: "bold")[双头架构]], [ (size: 11pt, weight: "bold")[物理 RL]], [ (size: 11pt)[随机交错CIF+物性+文本]], [ (size: 11pt)[语言推理+数值回归互增强]], [ (size: 11pt)[稳定性/新颖性/多样性协调]], ) ] ], ) Three-stage training (center)[ (size: head-s, weight: "bold", fill: accent)[三阶段分工明确：基座 → 预测 → 生成]] (0.2em) ( columns: (1.2fr, 1.5fr, 1.5fr, 2.2fr), inset: 4pt, stroke: 0.4pt, align: (left, center, center, left), table.header( [ (size: body-s, weight: "bold")[Stage]], [ (size: body-s, weight: "bold")[输入]], [ (size: body-s, weight: "bold")[输出]], [ (size: body-s, weight: "bold")[核心能力]], ), [ (size: 12pt)[1 基座构建]], [ (size: 12pt)[CIF+数值+文本]], [ (size: 12pt)[更好的 LLM]], [ (size: 12pt)[建立材料科学先验]], [ (size: 12pt)[2 SAR+双头]], [ (size: 12pt)[CIF + 问题]], [ (size: 12pt)[CoT推理 + 数值]], [ (size: 12pt)[正向：结构 → 性质]], [ (size: 12pt)[3 GRPO RL]], [ (size: 12pt)[指令]], [ (size: 12pt)[Wyckoff 序列]], [ (size: 12pt)[反向：性质 → 结构]], ) (0.3em) ( columns: (1fr, 1fr, 1fr), gutter: 0.6em, align: top, [ (title: "Stage 1: 随机交错预训练")[ (size: body-s)[ 基于 S1-Base 8B（Qwen3-8B）\ Alexandria 数据库稳定子集\ CIF + DFT 物性 + 文本随机交错\ 同一结构中三者轮流出现 \ → 跨模态关联自然建立 ] ] ], [ (title: "Stage 2: SAR + 双头预测")[ (size: body-s)[ 30K 专家标注 SAR 数据\ 语言头：CoT 推理，交叉熵 loss\ 回归头：MeanPool→Linear(1)，MSE\ Warm-up：先冻 backbone 训回归头→再联合 ] ] ], [ (title: "Stage 3: GRPO 物理 RL")[ (size: body-s)[ Wyckoff SFT → 语法合格\ GRPO：组内 G 个候选 → MLIP 弛豫\ → 4 种奖励 → 组内归一化 → 策略更新\ RL 贡献 S.U.N. +23.2pp ] ] ], ) Dual-head architecture -col( ratio: (1fr, 1fr), [ (title: "回归头设计")[ (size: body-s)[ 最后一层所有 token hidden state 做 mean pooling，经线性变换输出标量：\ hat(y) = W ("reg") dot 1/L sum (i=1)^L h i + b ("reg") \ W "reg" in bb(R)^(1 times d) 线性投影 \ MSE 损失： cal(L) "pre" = 1/N sum (hat(y) n - y n)^2 ] ] (title: "语言头（SAR SFT）")[ (size: body-s)[ 标准自回归 next-token prediction \ CoT 推理链：排序/区间判断/目标筛选 \ 损失 cal(L) "kno" = -1/T sum log p theta (x|x (<t)) ] ] ], [ (title: "联合损失 + Warm-up")[ (size: body-s)[ cal(L) ("all") = alpha cal(L) ("kno") + beta cal(L) ("pre") ] (size: body-s, weight: "bold")[ alpha , beta ：Pareto 前沿确定] (size: body-s, weight: "bold", fill: accent)[两步热身warm-up] (size: body-s)[ 回归头参数随机初始化，backbone 36 层已充分训练。直接联合导致随机梯度噪声破坏 backbone 表征。解决： 1. 冻结 backbone，只训回归头 \ 2. 解冻所有参数联合优化 ] (size: body-s)[ - cal(L) "pre" 梯度穿过backbone36层约束输出 \ - cal(L) "kno" 梯度约束语言推理 \ → hidden state 朝双向优化调整 ] ] ], ) Grpo multi-objective rl -col( ratio: (1fr, 1fr), [ (title: "GRPO 迭代流程")[ (size: body-s)[ 1. 策略 pi theta 生成 G 个 Wyckoff 序列 (e.g. G=12)\ 2. 每个序列 → CIF → MLIP 弛豫 → 算奖励 \ 3. 门控 V(C) ：间距≥0.5Å + 电中性 + 弛豫收敛 \ 4. 组内归一化： A i = frac(R(C i) - mu R, sigma R + epsilon) \ 5. 优势函数 A i 更新 theta ] ] (0.3em) (right)[ (size: 12pt)[ cal(L) ("GRPO") = - bb(E) [ 1/G sum (i=1)^G min( frac(pi theta(C i), pi ("old")(C i)) A i, "clip"( dots.c ) A i ) ] + beta dot D ("KL") ] ] ], [ (title: "4 种奖励")[ (size: body-s, weight: "bold")[ R S 稳定性] (size: body-s)[ 基于 E "hull" 凸衰减映射 \ E "hull" <= 0.1 arrow.r 1 \ E "hull" >= 0.2 arrow.r 0 \ 凸指数 gamma=2 聚焦近稳定结构\ ] (size: body-s, weight: "bold")[ R N 新颖性] (size: body-s)[StructureMatcher 硬门控 + 指纹距离软打分\ ] (size: body-s, weight: "bold")[ R U 多样性] (size: body-s)[组内指纹距离 → Shannon 熵 → 归一化\ ] (size: body-s, weight: "bold")[ R P 条件匹配] (size: body-s)[条件生成时属性一致性] ] ], ) Metrics -col( ratio: (1fr, 1fr), [ (title: "性质预测：MAE")[ (size: body-s, weight: "bold")[三个物理维度各异的属性] (0.2em) (size: body-s)[ (weight: "bold")[ E "hull" ] (eV/atom)：热力学稳定性 \ (weight: "bold")[体模量] (GPa)：力学功能属性 \ (weight: "bold")[带隙] (eV)：电子功能属性 ] ] (title: "性质预测结果")[ ( columns: (2.2fr, 1.3fr, 1fr, 1fr, 1.3fr), inset: 3pt, stroke: 0.4pt, align: (left, center, center, center, center), table.header( [ (size: 11pt, weight: "bold")[任务]], [ (size: 11pt, weight: "bold")[MatMind]], [ (size: 11pt, weight: "bold")[CGCNN]], [ (size: 11pt, weight: "bold")[M3GNet]], [ (size: 11pt, weight: "bold")[LLM-Prop]], ), [ (size: 11pt)[ E "hull" ]], [ (size: 11pt, fill: accent, weight: "bold")[0.0109]], [ (size: 11pt)[0.0123]], [ (size: 11pt)[0.0130]], [ (size: 11pt)[0.0138]], [ (size: 11pt)[体模量]], [ (size: 11pt, fill: accent, weight: "bold")[5.36]], [ (size: 11pt)[5.55]], [ (size: 11pt)[5.57]], [ (size: 11pt)[6.14]], [ (size: 11pt)[带隙]], [ (size: 11pt, fill: accent, weight: "bold")[0.197]], [ (size: 11pt)[0.244]], [ (size: 11pt)[0.209]], [ (size: 11pt)[0.346]], ) (0.2em) (size: 12pt, fill: secondary-text)[首次以 LLM 范式超越专用 GNN] ] ], [ (title: "晶体生成：S.U.N. 率")[ (size: body-s)[ 三重筛选的复合指标： ] (0.1em) (size: body-s)[ (weight: "bold")[S]table — E "hull" <= 0 ，热力学可合成 \ (weight: "bold")[U]nique — 组内 N 个结构两两不等价 \ (weight: "bold")[N]ovel — 不匹配 MP-20 + Alexandria 训练集 ] ] (0.3em) (title: "无条件生成 S.U.N.")[ ( columns: (2.5fr, 1.5fr), inset: 3pt, stroke: 0.4pt, align: (left, center), table.header( [ (size: 11pt, weight: "bold")[方法]], [ (size: 11pt, weight: "bold")[S.U.N. 率]], ), [ (size: 11pt, fill: accent, weight: "bold")[MatMind CIF RL]], [ (size: 11pt, fill: accent, weight: "bold")[65.3%]], [ (size: 11pt)[MatterGen]], [ (size: 11pt)[44.3%]], [ (size: 11pt)[DiffCSP]], [ (size: 11pt)[40.2%]], [ (size: 11pt)[MatMind SFT (无 RL)]], [ (size: 11pt)[42.1%]], ) (0.2em) (size: 12pt, fill: accent, weight: "bold")[RL 贡献 +23.2pp] (size: 12pt, fill: secondary-text)[，SFT 仅与扩散基线持平] ] ], ) Conditional generation and limitations -col( ratio: (1.2fr, 1fr), [ (title: "条件生成：极稀疏迁移")[ (size: body-s, weight: "bold")[输入条件：磁化密度 >= 0.2 mu B /Å^3 （正样本仅 (fill: accent, weight: "bold")[21 个]（60 万中））\ RL阶段前后的生成质量对比：\ Pre-RL 1.2% → Post-RL 5.2%\ ] (size: body-s, weight: "bold")[其他条件任务\ ] ( columns: (2fr, 1.5fr, 1.5fr), inset: 3pt, stroke: 0.4pt, align: (left, center, center), table.header( [ (size: 11pt, weight: "bold")[条件]], [ (size: 11pt, weight: "bold")[Pre-RL]], [ (size: 11pt, weight: "bold")[Post-RL]], ), [ (size: 11pt)[带隙 >5eV]], [ (size: 11pt)[18.0%]], [ (size: 11pt)[34.8%]], [ (size: 11pt)[体模量 300GPa]], [ (size: 11pt)[12.8%]], [ (size: 11pt)[26.8%]], ) 生成的 Gd₂FeIr 和 Gd₂MnCo₃ 在化学上合理 —— \ 模型学到了 (fill: accent, weight: "bold")[真正的构效关系]，而非插值。 ] ], [ (title: "局限性")[ (size: body-s, weight: "bold", fill: accent)[论文层面\ ] (size: body-s)[ · 参考文献大量 TODO xxx 占位符 \ · 消融实验未完成 \ · 注意力分析未完成\ ] (size: body-s, weight: "bold", fill: accent)[工程层面\ ] (size: body-s)[ · 8B 三阶段计算成本未披露 \ · 完全依赖 DFT/MLIP，未验证实验一致性 \ · 代码未开源，复现门槛高 \ · 依赖链长（NequIP / eSEN / MP 凸包） ] ] ], ) Q&A (4em) (center, text(size: 22pt, weight: "bold", fill: accent)[谢谢！Q & A]) (1em) (center, text(size: 12pt, fill: secondary-text)[arXiv: 2606.07712 · 代码暂未开源 · biscuit0613.github.io])