此次由中科大苏州高等研究院与北京深势科技有限公司联合举办的首届DeepModeling黑客马拉松大赛,希望通过汇集一批年轻的科研工作者、技术工程师交流开发经验、提升开发能力,为未来提供高效、可复用、可持续优化的建模工具探索新的协同模式。无论你是技术大神还是萌新小白,无论你是科研工作者还是技术工程师,你都可以尝试单独成队或和你的团队聚在一起实现你们的创意!无论你来自什么院校,来自什么部门,来参加就是了!
具体安排请关注“深度势能”公众号,并留意最新推送。
7月19日 ABACUS使用指导与赛题介绍
7月21日 DeePMD-kit使用指导与赛题介绍
7月23日 NVIDIA线下活动与直播教程大礼包(含NVIDIA、DeePMD-kit、PaddlePaddle关于科学计算中的高性能优化分享)
7月25日 科学计算工作流设计与案例介绍
7月26日 FEALPy使用指导与赛题介绍
7月28日 PaddlePaddle使用指导与赛题介绍
点击下载(code:683b)
https://hackathon.dp.tech/2021/summer-hackathon(请用电脑浏览器打开,暂不支持移动端显示)
7月1日:放出部分赛题,选手开始提交proposal及赛题意见
7月11日:更新最终赛题并确定数据集
7月12日:开始线上直播软件教学(具体直播安排见公众号推送)
7月30日:报名截止并停止提交proposal,开始对接导师指导交流
(7月1日—8月15日是大家熟悉题目、开始动手研究题目的好时机哦)
8月16日—17日:报告分享与教学:
聆听各位嘉宾的分享报告,学习使用相关软件工具,并随时交流答疑。
8月18日—20日:正式比赛:
放出bonus题目,各组选定题目比赛,赛后专家团队评分,选出优秀队伍做展示并颁奖。
每队7月30日之前可以反复提交proposal,于正式比赛当天确定1-2道题目,此部分题目的分数主要基于完成度和创新度;
Bonus题目可选多道,不需要提前选定,可根据时间安排即答即交,我们根据此部分题目的完成度和题目数量非线性给分。
1.Model interpretability is a crucial property for machine learning, which is also challenging for researchers.
2.Plenty of efforts have been made to develop Explainable Artificial Intelligence (XAI, see a review in here ), such as:
3.Descriptors in DeePMD-kit contain the environment information of each central atom, and are fed into fitting network to obtain the final output. The interpretability of these descriptors remains no more investigation, which is open to all kinds of responsible explanation.
1.Investigate the post-hoc interpretability of descriptors in copper systems of different crystal structures (i.e. descriptors of Cu
in different crystal structures) in DeePMD-kit.
2.Requirements:
(1)set descriptor
as se_a
and train 100w steps on 6 copper systems as a whole training set, then make analysis for 6 given systems (it will take hours to train, so it’s better to start up early);
(2)add type_one_side
= Fasle
of descriptor parameters in input.json, an example is provided here: input.json(code:fwpn), feel free to change other training parameters.
1.This is a problem for people who prefer to do some analysis and may not require much change to the code and implementation.
2.To get the basic score, you need to be familiar with the se_a
descriptor in deepmd-kit and the reasons for its construction. Design a method or utilize methods mentioned above to illustrate the interpretability of descriptors in DeePMD-kit, possibly do some clustering or visualization;
3.Feel free to do other relevant analysis and get the bonus score.
4.PaddlePaddle provides various tools convenient for analysis and Interpretation:
In addition, if you choose PaddlePaddle, mentorship can be provided.
You can see here for coding instruction, and hackathon2021
branch is where you accomplish this project. You might be able to get the descriptors mainly around deepmd-kit/deepmd/descriptor
in hackathon2021
branch. Be aware of online tutorial of DeePMD-kit and online tutorial of PaddlePaddle, which are important for those who choose DeePMD-kit relevant projects.
1.Copper systems in different crystal structures: Cu_full.zip(code:cwur).
1.The rationality and completeness of the analysis.
2.The innovativeness of the analysis.
A zip file which contains:
1.a report detailing the process of the experiment and the analysis results.
1.Model compression in AI is solution to efficiency and information distillation.
2.Networks in DeePMD-kit may contain some redundant parameters.
3.The information quantity contained in different layers is not clearly compared and may be a way to do network pruning.
1.Train the DeePMD-kit 100w steps on the given system below and freeze the model (feel free to set other parameters in input.json);
2.Make an analysis on the information quantity contained in different layers;
3.Use some strategies to do model compression and make the network smaller and more efficient without losing too much accuracy.
1.To get the basic score, you need to have a clear understanding of the overall model structure of DeePMD-kit.
2.You can use mutual information to measure the information quantity contained in different layers;
3.You can use quantization、knowledge distillation、low-rank factorization or PCA and so on to do model compression;
4.PaddlePaddle provides various tools convenient for analysis and model compression:
In addition, if you choose PaddlePaddle, mentorship can be provided.
You can see here for coding instruction, and hackathon2021
branch is where you accomplish this project. You might be able to get the descriptors mainly around deepmd-kit/deepmd/descriptor
in hackathon2021
branch. Be aware of online tutorial of DeePMD-kit and online tutorial of PaddlePaddle, which are important for those who choose DeePMD-kit relevant projects.
1.Small dataset: water example contained in example/water/data
, you can also download from here;
2.Challenge dataset: Cu system: cu.hcp.02x02x02.zip(code:4c21) (which is one of the six systems of dataset in A1), and an example input.json is provided here: input.json(code:fwpn)(which is the same as in A1).
1.The rationality and completeness of the analysis.
2.The effectiveness and efficiency of the compression (the amount of parameters in compressed model, the inference speed and so on, compared with the standard model).
A zip file which contains:
1.a report detailing the process of the experiment and the analysis results.
2.a copy of code that can run directly (a trained model included).
1.Neural Architecture Search(NAS) may be solution to tedious network architecture design(https://arxiv.org/abs/1707.07012), which can learn the model architectures directly on the dataset of interest.
2.Architectures in DeePMD-kit are delicately designed based on linear and resnet blocks with theoretical hypothesis,while we are open to any other available architecture design.
1.Use NAS(or meta learning) to automatically change parameters (neuron
, lr
and pref
are recommended) in input.json for better performance in energy & force fitting, after 20w steps training on the data systems given below;
2.Use NAS to search for a better network, which may outperform standard DeePMD-kit, after 100w steps training on the data systems given below.
1.This problem is the one with high freedom, you may use any tools to achieve the goals.
2.For Goal2, in energy/force fitting (or you can simultaneously do both), you can use NAS to search for the best architectural building block on a small part of dataset and then transfer the best block architecture to a larger part, then train a new model. Finally, compare with standard trained DeePMD-kit on validation part.
3.PaddlePaddle provides various tools convenient for NAS:
In addition, if you choose PaddlePaddle, mentorship can be provided.
You can see here for coding instruction, and hackathon2021
branch is where you accomplish this project. You might be able to get the descriptors mainly around deepmd-kit/deepmd/descriptor
in hackathon2021
branch. Be aware of online tutorial of DeePMD-kit and online tutorial of PaddlePaddle, which are important for those who choose DeePMD-kit relevant projects.
1.Small dataset: water example contained in example/water/data
, you can also download from here;
2.Challenge dataset: Cu system: cu.hcp.02x02x02.zip(code:4c21) (which is one of the six systems of dataset in A1), and an example input.json is provided here: input.json(code:fwpn) (which is the same as in A1).
1.The effectiveness and correctness of the NAS procedure;
2.The innovativeness of the NAS procedure.
A zip file which contains:
1.a report detailing the process of the experiment;
2.a copy of code that can run directly (a trained model included).
1.Mixed precision training is widely used in HPC. Under the premise of ensuring the output accuracy within a certain range, the training process can be accelerated by using single precision or semi-precision.
2.DeePMD-kit adopts double precision training by default and has a single precision training interface, but there is no corresponding exploration for semi-precision training.
1.With single v100 and specific Tensorflow version (will be given later), try to use the mixture of single precision and semi-precision in the DeePMD-Kit training process, and speed up the training process under the premise of ensuring the output precision (threshold of error will be given later).
2.After the same time training process (you can set the time), compare the validate loss with the standard training procedure.
1.You can feel free to change the float precision during training process;
2.To achieve the best performance, you can also change the neuron parameters to refine the number of network layers;
3.Note:To ensure the accuracy of semi-precision training, attention should be paid to gradient explosion and gradient vanishing. In addition, be careful about the matrix dimension in semi-precision Tensor Core.
You can see here for coding instruction, and hackathon2021
branch is where you accomplish this project. You might be able to get the descriptors mainly around deepmd-kit/deepmd/descriptor
in hackathon2021
branch. Be aware of online tutorial of DeePMD-kit, which is important for those who choose DeePMD-kit relevant projects.
1.Small dataset: water example contained in example/water/data
, you can also download from here;
2.Challenge dataset: Cu system: cu.hcp.02x02x02.zip(code:4c21) (which is one of the six systems of dataset in A1), and an example input.json is provided here: input.json(code:fwpn) (which is the same as in A1).
1.The correctness and efficiency of the training boosting ;
2.The innovativeness of the procedure.
A zip file which contains:
1.a report detailing the process of the experiment;
2.a copy of code that can run directly (a trained model included).
1.For large-scale DFT calculations, ABACUS adopts atomic orbitals with strict radius cutoffs. Efficient searching and recording neighboring atoms is an essential procedure that has been implemented in ABACUS. The current algorithms in searching neighbors can be used for large systems, but still has potentials to be further improved.
1.Read and understand the algorithms for searching neighboring atoms.
2.Remove the dependence of the Hash library in BOOST math library by implementing your own algorithm or using the STL in C++. Ensure the new algorithm can work correctly and compare the efficiency of the new and old algorithms.
3.Try to improve the efficiency of searching neighbors.
1.Use MPI, openMP or CUDA to accelerate the efficiency of the code.
You can see here for coding instruction, and hackathon2021
branch is where you accomplish this project. The code to be edited might mainly lies in abacus-develop/source/module_neighbor
in hackathon2021
branch. Be aware of online tutorial of ABACUS, which is important for those who choose ABACUS relevant projects.
1.Small dataset: example dataset in code package: abacus-develop/tests/501_NO_neighboring_GaAs512/
, you can also download here: 501_NO_neighboring_GaAs512.zip(code:8vvm).
2.Challenge dataset: C262144-neighbor, download here: C262144-neighbor.zip(code:n0q1).
1.The correctness of the implementation;
2.The efficiency of the code compared with the standard ABACUS procedure.
A zip file which contains:
1.a report detailing the process of the experiment;
2.a copy of code that can run directly.
1.Assembling the element matrix is a key step in the finite element method. To get the element matrices, one need to calculate the numerical integral of the product between any two local basis functions on each mesh element.
2.At present, FEALPy uses the einsum function in Numpy to get the element matrices,but this function does not support multi-core calculation. As a result, FEALPy can not make full use of the computer’s multi-core computing resources in this step.
Please see the code in construct_stiff_matrix.py(code:07q1), design and develop a special multi-core version function to replace the einsum function.
One can find some detailed discussion in the following paper:
Luporini F, Varbanescu A L, Rathgeber F, et al. Cross-loop optimization of arithmetic intensity for finite element local assembly[J]. ACM Transactions on Architecture and Code Optimization (TACO), 2015, 11(4): 1-25.
Please see the install.md(code:42ho) file for installation FEALPy on your system. See here for coding instruction, and hackathon2021
branch is where you accomplish this project. Be aware of online tutorial of FEALPy, which is important for those who choose FEALPy relevant projects.
1.The correctness of the implementation;
2.The efficiency of the code compared with the standard single-core FEALPy procedure.
A zip file which contains:
1.a report detailing the process of the experiment;
2.a copy of code that can run directly.
1.Well-designed workflows are important for the transparency and reproducibility of scientific computing tasks. In addition, they are very useful for both pedagogical and production purposes.
2.Practitioners in scientific computing typically lack trainings for managing and maintaining workflows.
3.Here we list a few tasks for which we pay particular attention to the workflow perspective. One may propose their own workflows(choose one or more of the followings, or you can inform us and put forward other workflows):
-
to compute the heat conductance of water using a Deep Potential model;
-
to compute the radial distribution functions using a Deep Potential model;
-
to compute the diffusion coefficient using a Deep Potential model.
Develop good workflows for large-scale and computationally-intensive tasks, which would help to boost the efficiency of scientific computation jobs.
Design and develop a workflow using Apache airflow or aiida, or other workflow management tools. One may take dpti (https://github.com/deepmodeling/dpti ) as an example.
You can see here for coding instruction, and hackathon2021
branch is where you accomplish this project. You might be able to get the descriptors mainly around deepmd-kit/deepmd/descriptor
in hackathon2021
branch. Be aware of online tutorial of DeePMD-kit and online tutorial of Workflow, which are important for those who choose DeePMD-kit relevant projects.
1.The correctness of the implementation;
2.The universality and transferability of the designed workflow.
A zip file which contains:
1.a report detailing the process of the experiment;
2.a copy of code that can run directly.
1.ABACUS is a density functional theory software based on quantum mechanics, it can be used to predict properties of materials. It is a powerful tool to combine ABACUS with open database for materials, such as Materials Project.
1.Developing a workflow to calculate band gaps of materials with ABACUS for at least 100 structures downloaded from the Materials Project. (Compounds of III-V family elements and II-VI family elements are recommended.)
2.Make sure that the workflow code has good universality for other materials.
You can see here for coding instruction, and hackathon2021
branch is where you accomplish this project. Be aware of online tutorial of ABACUS, which is important for those who choose ABACUS relevant projects.
1.An example including input files and explanations for key parameters is provided: example-Si-band.zip(code:i2fs) , which contains the first scf step and the second nscf step of Si-diamond band calculation.
2.The atomic structures and related information can be downloaded from the Materials Project, and there is an official API to get data from this database.
3.You can find similar workflow examples on this website.
4.You can download pseudopotential files here.
5.The goal 100 material examples are listed in Material_IDs.txt(code:j1d9), which you need to download from Materials Project and compute with ABACUS.
1.The correctness of the implementation;
2.The universality and transferability of the designed workflow.
A zip file which contains:
1.a report detailing the process of the experiment;
2.a copy of code that can run directly.
1.In FEALPy, the basis functions of Lagrange finite element space defined on the simplex (interval, triangle or tetrahedron) meshes is constructed based on the barycentric coordinates, which does not need to introduce the reference element. Furthermore, they satisfy the interpolation property, that is, each basis function takes 1 at one of interpolation points and 0 at the other on each element.
2.The calculation of higher derivatives and numerical integrals of this kind of basis function is cumbersome.
3.If the interpolation property is not required, one can use the Bernstein polynomial based on barycentric coordinates, which owns simpler form and has clear formulas for derivation and integral calculation.
1.Design and develop a space class named BernsteinFiniteElementSpace
, which should have same interfaces as the LagrangeFiniteElementSpace
class in FEALPy;
2.All function should be implemented by the array-oriented and dimension independent techniques;
3.Verify the correctness of this space class by solving Poisson equation.
1.One can find the definition of Bernstein polynomial in Wiki.
2.One can find the Bernstein basis based on barycentric coordinates in the follow paper:
Feng L, Alliez P, Busé L, et al. Curved optimal delaunay triangulation[J]. ACM Transactions on Graphics, 2018, 37(4): 16.
3.Please see S3_implementation(code:dkgl) and download files for the specific code implementation requirements.
Please see the install.md(code:42ho) file for installation FEALPy on your system. See here for coding instruction, and hackathon2021
branch is where you accomplish this project. Be aware of online tutorial of FEALPy, which is important for those who choose FEALPy relevant projects.
1.The correctness of the implementation;
A zip file which contains:
1.a report detailing the process of the experiment;
2.a copy of code that can run directly.
DeePMD-kit is a package written in Python/C++, designed to minimize the effort required to build deep learning based model of interatomic potential energy and force field and to perform molecular dynamics (MD). This brings new hopes to addressing the accuracy-versus-efficiency dilemma in molecular simulations.
Find better training parameters and strategies than the given input script.
1.Follow the tutorial and train a model for only 10w steps to predict the energy、force of given systems.
2.Change training parameters or strategies in the input.json such that the learning curve in lcurve.out drops faster than that in the original one, especially the training loss in the first 1w steps.
3.Use some plot tools to show the difference before and after the change.
4.You can choose DeepMD-kit based on Tensorflow (B1) or based on PadddlePaddle (B2).
You can see here for coding instruction, and hackathon2021
branch is where you accomplish this project. You might be able to get the descriptors mainly around deepmd-kit/deepmd/descriptor
in hackathon2021
branch. Be aware of online tutorial of DeePMD-kit and online tutorial of PaddlePaddle, which are important for those who choose DeePMD-kit relevant projects.
Dataset: Cu system: cu.hcp.02x02x02.zip(code:4c21) (which is one of the six systems of dataset in A1), and an example input.json is provided here: input.json(code:fwpn)(which is the same as in A1).
1.The correctness of the implementation;
A zip file which contains:
1.a report showing the results with a brief analysis.
2.input scripts and trained models both before and after the change.
ABACUS is an electronic structure package based on density functional theory (DFT). ABACUS adopts either plane wave basis or numerical atomic orbitals.
Please complete 2 of 4 calculations in the following designated physical properties with ABACUS software:
1.Draw electronic bands of GaAs crystal. Do self-consistent calculation first and do non-self-consistent calculation next with high symmetry K line.
2.Find bond length of N dimer. Spin polarization should be considered, and using two N atoms in large cell to simulate dimer, you can use energy fitting or atomic relaxation method.
3.Do molecular dynamics NVT simulation of Sn64 system. Draw pair distribution function to verify your results.
4.Calculate the optimal structure of Zn crystal with ABACUS. You can use cell relaxation method.
You can design input parameters in INPUT, STRU and KPT files by yourself with provided pseudopotential and numerical orbital files if needed. Plane wave base is recommended but feel free to try numerical orbital base.
You can see here for coding instruction, and hackathon2021
branch is where you accomplish this project. Be aware of online tutorial of ABACUS, which is important for those who choose ABACUS relevant projects.
Inputs of above 4 designated physical properties can be found in ABACUS_B3(code:vkhh).
1.The correctness of the implementation;
A zip file which contains:
1.a report showing the results with a brief analysis.
2.a copy of ABACUS input files.
1.FEALPy is a Python package to numerically solve PDEs, but the example demos in FEALPy is far from comprehensive.
2.One can learn how to use FEALPy by reading and running the demo scripts in the FEALPy/example
directory, such as:
PoissonFEMWithDirichletBC_example.py
PoissonFEMWithNeumannBC_example.py
PoissonFEMWithRobinBC_example.py
ConvectinDiffusionReactionFEMwithDirichletBC2d_example.py
Add more examples into FEALPy.
There are many demos in FEniCS Project, such as
(1)Poisson equation with multiple subdomains
(2)Cahn-Hilliard equation
Try to implement one of the above two examples based on FEALPy.
Please see the install.md(code:42ho) file for installation FEALPy on your system. See here for coding instruction, and hackathon2021
branch is where you accomplish this project. Be aware of online tutorial of FEALPy, which is important for those who choose FEALPy relevant projects. Scoring point:
1.The correctness of the implementation;
A zip file which contains:
1.a report briefly introducing how you made it.
2.a copy of code file such as PoissonEquationWithMultipleSubdomains_example.py or CahnHilliardEquation_example.py, which can run directly.