岗位描述
我们在建设面向化学与材料研发的生成式 AI 能力，目标是把生成模型与评估/模拟/实验数据协同起来，形成可迭代、可交付的研发工作流。你将负责从数据到模型到决策层的关键模块研发，并与领域专家一起把科学约束变成可计算的模型目标。工作职责（Responsibilities）：你将围绕“数据—生成—评估—迭代”的核心链路开展工作：数据与表示：整合公共数据库与内部数据（实验/计算/文献），完成清洗、标准化、去重与数据质量评估，构建可用于训练与评测的高质量数据集。生成模型研发：设计、训练与优化生成式模型（如 Transformer/GPT、Diffusion、Flow Matching 等），支持面向化学/材料对象的结构生成与优化。多目标与约束生成：将多种性能目标与物理/化学约束转化为可计算的条件，实现可控生成与多指标权衡。评测与决策层：建立可靠的评测体系与决策层（打分/过滤/重排序/不确定性），提升结果的有效性、鲁棒性与可复现性。工程化与协作交付：与领域专家及实验/模拟团队协作，搭建可复现 pipeline（训练/推理/评测/监控），输出可执行候选与阶段性报告，支持闭环迭代。任职要求（Requirements）“硕士及以上学历（博士优先），计算机/人工智能/计算化学/材料科学等相关方向。扎实的深度学习能力，熟练使用 PyTorch/TensorFlow，能编写整洁、模块化、可维护的研究/工程代码。具备生成式模型实战经验：至少熟悉并用过 Transformer/GPT、Diffusion、Flow Matching 中的一类（训练、调参、推理、评测）。良好的数据与系统能力：能独立完成数据管线、实验管理、指标设计与误差分析。跨学科协作能力强：能把领域约束抽象成算法问题，并推动落地。加分项（Preferred）：AI for Science 相关落地经验（化学/材料/生物/物理）。熟悉 RDKit / 分子与材料结构处理工具链 / 图神经网络（GNN）/ 强化学习（RL）。熟悉分布式训练、加速推理、工业级调参与实验追踪（如 wandb/mlflow/optuna 等）。有将模型接入真实研发流程（服务化、稳定性、监控、回归测试）的经验We are building generative AI capabilities for chemistry and materials R&D. Our goal is to integrate generative models with evaluation/simulation/experimental data to form an iterative, deliverable R&D workflow. You will develop key modules across the stack—from data to models to decision-making layers—and work closely with domain experts to turn scientific constraints into computable optimization objectives.Responsibilities：You will work end-to-end along the core loop: Data → Generation → Evaluation → Iteration:Data & Representations：Integrate public datasets and internal data sources (experiments, computations, literature). Perform cleaning, standardization, deduplication, and data quality assessment. Build high-quality datasets for training and benchmarking.Generative Model Development：Design, train, and optimize generative model (e.g., Transformers/GPT, Diffusion, Flow Matching). Support structure generation and optimization for chemistry/materials objects.Multi-objective & Constrained Generation：Translate multiple property targets and physical/chemical constraints into computable conditions. Enable controllable generation and trade-off optimization across multiple metrics. Evaluation & Decision Layer：Build reliable evaluation and decision modules (scoring, filtering, re-ranking, uncertainty estimation). Improve validity, robustness, and reproducibility of generated results.Engineering & Cross-functional Delivery：Collaborate with domain experts and experimental/simulation teams. Build reproducible pipelines (training/inference/evaluation/monitoring). Deliver actionable candidates and milestone reports to support closed-loop iteration.Requirements：M.S. degree or above (Ph.D. preferred) in CS/ML/computational chemistry/materials science or related fields. Strong deep learning fundamentals; proficiency with PyTorch/TensorFlow; ability to write clean, modular, maintainable research/production code.Hands-on experience with at least one class of generative models: Transformer/GPT, Diffusion, or Flow Matching (training, tuning, inference, evaluation).Strong data/system skills: able to independently build data pipelines, manage experiments, design metrics, and perform error analysis.Strong cross-disciplinary collaboration skills: able to abstract domain constraints into algorithmic problems and drive implementation.Preferred Qualifications：· Applied experience in AI for Science (chemistry/materials/biology/physics).Familiarity with RDKit, molecular/materials structure toolchains, GNNs, and/or reinforcement learning. Experience with distributed training, inference acceleration, and industrial-grade tuning/experiment tracking (e.g., Weights & Biases, MLflow, Optuna).Experience integrating models into real R&D workflows (service deployment, reliability, monitoring, regression testing).