jse-release-v4.1.2

🚀 Features

• Added setLossOut(String path) to TrainerNNAP for convenient real-time loss output, along with support for retrieving sub-losses.
• TrainerNNAP now supports disabling neighbor caching (experimental).
• NativeLmp now provides direct access to commRank(), commSize(), and commBarrier().

⚡ Performance Improvements

• Optimized NNAP GPU performance: different cutoff radii now use separate neighbor lists.
• TrainerNNAP automatically disables forward propagation caching for large atomic structures, improving parallel training speed.
• TrainerNNAP merges neighbor list pointer storage, significantly reducing memory usage.
• Optimized type checking for method calls in SimpleJIT, improving performance for large-scale invocations.

✨ Enhancements

• The runtime now proactively detects and distinguishes between JRE and JDK, and throws an exception when running on an unsupported JRE.
• parfor now strictly validates the thread count to prevent undefined behavior.
• Added parallel initialization detection to avoid catastrophic failures.
• Stricter type checking in DataXYZ.setParameter to prevent common misuse.

🐛 Bug Fixes

• Fixed redundant duplicate error messages when running jse -lmp in parallel.
• Fixed a NullPointerException in Lmpdat.of when data without mass is provided.

🧹 Removals

• Removed the deprecated doc_agent. Related content should now be referenced at jse-skil.

---

Note
For other potentially critical updates, see v4.1.0.

jse-release-v4.1.1

Zero length cptr fill and other bug fixes from v4.1.0

jse-release-v4.1.0

Changelog

Breaking Changes

• JIT method invocation API changed: IJITMethod.invoke(IPointer, IPointer) is now invoke(Object...). You no longer need to manually construct mDataIn/mDataOut parameter buffers — just pass arguments directly.
• Trainer renamed to TrainerNNAP: some constructor parameter signatures have also changed.
• WTypeBasis model format changed: the post_fuse mechanism has been replaced by rfuse. Legacy models are automatically converted on load. The default training wtype changed from exfull to rfuse.
• Growable CPointer variants removed: GrowableDoubleCPointer, GrowableFloatCPointer, GrowableIntCPointer, GrowableInt64CPointer, and their GPU counterparts have been removed. Use PointerManager for unified management instead.

New Features

• PointerManager: unified lifecycle management for CPointer / CudaPointer, implementing AutoCloseable with automatic reclamation via Java GC. NEP / NNAP internals have been fully migrated; close() is now significantly simpler.
• JIT SourceScanner: automatically discovers exported functions in C source code via the __jsefunc__ marker, eliminating the need for manual setMethodNames calls. SimpleJIT.Engine also gains addTypeMap for C type mapping.
• NNAP GPU support: PairNNAP_gpu is re-enabled with major performance improvements.
• Adam AMSGrad: the Adam constructor now accepts a boolean aAMSGrad parameter, enabling the improved algorithm from "On the Convergence of Adam and Beyond".
• Added SettableAtomData.setSymbolOrder(): element symbol order modification is now supported consistently across all atom structure classes.
• LibVer: centralized external library version management (JEP, LAMMPS, CMake, MiMalloc, Ninja, NEP SHA).

Bug Fixes

• FixedTimer: getNanos() / getMillis() / get() incorrectly returned the fixed value captured at construction time. They now correctly compute the actual elapsed time.
• AseAtoms.copy(): when mType2AtomicNumber != null, the symbol collection loop missed the last element type (type < ntypes → type <= ntypes), causing the copy to fail.

Improvements

• Naming consistency: Matrices parameters aRowNum/aColNum → aNumRows/aNumCols; SettableAtomData field mAtomTypeNum → mNumTypes; PairNEP / PairNNAP field mTypeNum → mNumTypes.
• Build system: Windows build scripts now include chcp 65001; CMake path references uniformly use JNIUtil.validCmdPath().
• Unified SF and RDF behavior in APC: both SF and RDF descriptor lengths now consistently use the input N, removing the previous N+1 discrepancy.

Deprecations

• SP.Python.downloadPackage() / installPackage() methods are now marked @ApiStatus.Obsolete. installAse() has been removed.

Internal

• Native static methods in CPointer subclasses now have @ApiStatus.Internal public visibility.
• ICPointer / ICudaPointer now expose a setPtr_() method.
• IO.Text.str2int() removed; call sites now use Integer.parseInt().
• External LAMMPS library updated from stable_22Jul2025_update2 to stable_22Jul2025_update4.

jse-release-v4.0.1

Changelog

Experimental Runtime Compilation (JIT)

• Added experimental runtime C/C++ compilation support (SimpleJIT).
• Added experimental CUDA runtime compilation support (CudaJIT).
• NNAP and NEP are now implemented via code generation + runtime compilation, delivering at least a 20% performance improvement and precision switching support.

CUDA Support

• Introduced basic CUDA memory management: CudaCore, FloatCudaPointer, IntCudaPointer.

Build System

• The build process has been unified to use CMake + Ninja.

Native LAMMPS Interface

• Added closeAll(), closePython(), closePlugin() (and similar) to NativeLmp, matching the LAMMPS C interface.
• All native interface commands now consistently end with 0; commands with the same base name receive an incremented numeric suffix.
• By default, errors from LAMMPS calling jse are now only thrown on rank 0, preventing excessive duplicate error messages.

I/O, Data Formats & Strict Mode

• Added XML read/write support (IO.xml2map, IO.map2xml).
• Renamed confusing properties such as atomNumber → natoms; removed the old names.
• XDATCAR now supports reading files with per-frame cell information, and writes always use this format.
• Streaming read/write interfaces for atomic data are now public; the underscore suffix has been removed.
• Added step(int) to multi-frame trajectory readers (Lammpstrj, XDATCAR, DumpXYZ) for interval-based dumping.
• Reading of atomic structure data now defaults to strict mode: errors throw exceptions instead of returning null.
• Because reads are strict, streaming consumers should now use FileEndException to distinguish end-of-file from other errors.

ASE Integration

• Completed property support for AseAtoms (info, array, calc.results); conversion to/from ASE atoms is now essentially lossless.
• Added special-case handling when converting between DataXYZ and AseAtoms to preserve as much data as possible (ASE standard).
• Unified ASE support under the ase() method; atomic data can be converted to ASE Atoms (PyObject) via ase(). asAseCalculator() is deprecated and kept temporarily (emits a deprecation warning).
• Atomic data now uniformly provides setSymbolOrder(String...) to adjust the mapping between element symbols and type indices—particularly useful for LAMMPS data files.

NNAP Refinements

• Reordered summation: changed from post_fuse to an rfuse-style summation, which together with exfuse significantly improves efficiency.
• Optimized force backpropagation: RnGrad is computed directly instead of RnPxyz, then gradnlj and finally graddxyz are derived in a unified manner.
• Shared neural networks in NNAP no longer use a separate C implementation; they now reuse the non-shared implementation path.
• Removed the noradial parameter.
• The new NNAP JIT implementation supports second-order backpropagation, rewriting the NNAP trainer workflow.
• The new NNAP trainer supports missing energy/force/stress data and parallel neighbor-list construction.
• The new trainer no longer supports disabling basis-parameter training; alternative mechanisms will be introduced later.
• The new trainer enables post_fuse and exfull by default.
• Added non-dimensionalization of the distance unit in the new trainer, where the unit length is obtained from the neighbor list to avoid artifacts from gas-phase configurations.

Plugin Per-Atom Output

• LmpPlugin now supports per-atom output (setPeratomFlag, setSizePeratomCols, setPeratomFreq, setVectorAtom, setArrayAtom).

Data Structures & Math Changes

• Added single-precision floating-point data structures: FloatList, FloatVector, FloatCPointer.
• Added LogicalMatrix.
• Removed square-bracket indexing for matrices.
• Merged ShiftVector (and similar) functionality into the standard Vector; RowMatrix now also supports shift operations.

Cleanup & Deprecations

• Removed the overused shutdown() method; the standard close() interface is now used everywhere.
• The --jniclean flag now only clears the cache by default; use --jniclean all to clean all libraries.
• Cleaned up deprecated Torch-related code.
• Removed deprecated opt() compatibility; only op() remains.

Bug Fixes

• Fixed an issue where jse downloads ignored proxy settings; now respects system proxy and the HTTPS_PROXY environment variable.

jse-release-v3.99.1b

This is a pre-release version for the nnap GPU version.

Use it with:

pair_style   jse jsex.nnap.PairNNAP_gpu
pair_coeff   * * path/to/nnpot.json Cu Zr

jse-release-v3.13.7

JNI auto-install and other detail fixes from v3.13.6

jse-release-v3.13.6

Deadly bug and other detail fixes from v3.13.5

jse-release-v3.13.5

• Added support for the latest JDK 25 and LAMMPS 2025.
• NNAP now provides a pure Java implementation of the basis, eliminating the need to go through JNI when performing basis calculations in APC.
• The NNAP Trainer now disables L2 regularization by default due to its limited effectiveness.
• Removed a large number of NNAP features that showed limited practical benefit but significantly increased code complexity.
• Added support for configuring CMAKE_SETTING via environment variables.
• The uniqueID has been uniformly expanded to a 16-character length.
• Added support for utilities such as json2list, list2json, and related features.
• Significantly optimized the initialization of JNI-related libraries, making the process smarter and more automated.
• Refactored the jse.lmp.NativeLmp structure: core functionality has been moved to jse.clib.LmpCore, with complete version isolation.
• Added support for running the internal LAMMPS directly via jse -lmp.
• Introduced new commands jse --jniclean and jse --jnibuild.
• Standardized command formats.
• Provided a fully automated installation script and updated the related documentation.

jse-release-v3.12.6

• Added warnings for CRT conflicts in JNI libraries on Windows
• Unified all NNAP gradient computations under barkward method
• Refactored the NNAP JNI layer to consistently use C++ templates
• Pre-sorted by type when handling multiple type struture in NNAP to improve cache hit rate
• Added support for fuse and exfuse type weights in NNAP
• Added l4max (≤ 3), extended l3max to 6, and extended lmax to 12 in NNAP
• Added support for continued training and the Adam optimizer in NNAP
• Added support for shared FFNNs in NNAP
• Added post_fuse support in NNAP
• Renamed package opt to optim
• Added velocities and masses interfaces in LMP plugin
• Fixed an issue where native LMP did not call MPI_Comm_f2c when obtaining comm
• The MPI JNI interface now enables Java array copying by default to ensure compatibility

jse-release-v3.10.0

• LmpPlugin 增加调用 ForwardComm 和 ReverseComm 通讯支持
• MPI 接口现在默认不再进行数组拷贝，性能提高
• MPI 接口增加数组越界检查
• JNI 库初始化会增加更多检测和提示
• 更新 groovy，gradle，mimalloc，jep，jsch 库，现在应该支持各种先进的 ssh 加密算法
• 增加通用的优化器支持（GD，CG，Adam，LBFGS）
• 纯 java 的 NNAP 训练器支持，现在不再依赖 torch，且效率更高
• 部分数学运算增加 C 实现可选（sum，prod，dot，norm1，matmul）
• soft 势支持
• EAM 势支持
• toml 格式读写支持
• 各种 bug 修复