Bridging Visual Intuition and Chemical Expertise: A Virtual Collaborative Framework for Autonomous Nonadiabatic Dynamics Analysis

Abstract

Analyzing nonadiabatic molecular dynamics trajectories traditionally heavily relies on expert intuition and visual pattern recognition, a process that is difficult to formalize. We present VisU, a vision-driven framework that leverages the complementary strengths of two state-of-the-art large language models (Doubao-Seed-1.6-Vision and DeepSeek-V3.2) to establish a “virtual research cooperation”. This operates through a “Mentor–Engineer–Student” paradigm that mimics the collaborative intelligence of a professional chemistry laboratory. Within this ecosystem, the Mentor provides physical intuition through visual reasoning, while the Engineer adaptively constructs analysis scripts, and the Student executes the pipeline and manages the data and results. VisU autonomously orchestrates a four-stage workflow comprising Preprocessing, Recursive Channel Discovery, Important-Motion Identification, and Validation/Summary. This systematic approach identifies reaction channels and key nuclear motions while generating a professional academic report at the end. By bridging visual insight with chemical expertise, VisU establishes a new framework for human–AI collaboration in the analysis of excited-state dynamics simulation results, significantly reducing dependence on manual interpretation and enabling more intuitive, scalable mechanistic discovery in nonadiabatic dynamics simulations.

Introduction

Photoinduced reactions are ubiquitous in photophysics, photochemistry, and photobiology. Understanding excited-state processes is therefore of central importance. Unlike ground-state reactions, where molecules typically relax on a single potential energy surface, excited-state dynamics often encounter regions where two or more electronic states become degenerate (conical intersections). In these regions, the Born–Oppenheimer approximation breaks down, and nuclear motion cannot be treated independently of electronic motiona regime known as nonadiabatic dynamics. − Nonadiabatic effects are widespread in polyatomic systems and govern many key photochemical events, such as internal conversion, intersystem crossing, and electron transfer. Studying nonadiabatic dynamics is therefore essential for predicting excited-state lifetimes, reaction quantum yields, and branching ratios. However, simulating these processes faces major challenges: the strong coupling between electrons and nuclei, the large number of degrees of freedom, and the need to treat multiple electronic states simultaneously. ,

A practical and widely used method for simulating nonadiabatic dynamics is trajectory surface hopping (TSH). In this framework, nuclei move classically on a single potential energy surface and may “hop” to another surface with a probability determined by the electronic wave function. This mixed quantum-classical approach offers two key advantages: it can describe arbitrary and anharmonic nuclear motions without predefining reaction coordinates, and it is computationally efficient enough to treat polyatomic systems at full dimensionality. − As a result, TSH can reveal atomic-level, real-time molecular motions during photoinduced reactions, such as bond breaking, ring puckering, and hydrogen transfer. For polyatomic systems, trajectory-based nonadiabatic molecular dynamics (NAMD) has become a standard tool because it balances computational cost with the ability to capture complex photoreaction mechanisms. ,,,

The analysis of high-dimensional NAMD trajectories, particularly in multichannel scenarios, presents a fundamental bottleneck. , The composite nature of molecular motions makes subtle branching behaviors difficult to identify without intensive, expert-guided interpretation. The difficulty lies in how chemists actually interpret trajectory data. Distilling chemically meaningful insights from NAMD outputs requires both visual perception and domain-specific reasoning. By visual perception, we mean the ability to inspect trajectory animations, molecular geometries, and structural changes over time, for example, following the evolution of bond lengths, angles, or dihedral angles, or visually comparing different molecular conformations. By reasoning, we mean drawing conclusions from what is seen based on chemical knowledge, for instance, identifying which atomic motions are likely to drive the system toward a conical intersection, or recognizing whether a particular structural deformation is chemically relevant. By domain-specific knowledge, we refer to an understanding of nonadiabatic dynamics (e.g., conical intersections, hopping mechanisms, reaction coordinates) and familiarity with all analysis approaches (e.g., plotting of time-dependent features, employing machine learning and so on).

In a traditional manual analysis, a researcher routinely performs this loop: watching trajectory animations, examining critical structures, and correlating molecular spatial patterns with photochemical concepts. This tightly coupled visual-chemical reasoning loop is essential for distinguishing nonadiabatic channels and identifying governing nuclear motions. However, this manual approach faces severe limitations. For complex molecules with many atoms and multiple coupled motions, the relevant patterns are often hidden in high-dimensional data. A researcher must examine numerous trajectories, test many possible internal coordinates, and repeatedly refine her interpretation. This process is time-consuming and heavily dependent on individual expertise.

Recent advances in unsupervised machine learning (ML) provided valuable tools to address this bottleneck. Dimensionality reduction techniques such as principal component analysis (PCA) , can project high-dimensional geometric data into low-dimensional spaces, and clustering algorithms can automatically group similar hopping geometries. These methods simplify the data and can reveal patterns that are not obvious from raw trajectories. Indeed, a growing body of work has applied unsupervised ML to NAMD analysis. ,− Nevertheless, these purely data-driven methods still lack the interpretive capacity of a chemist. For instance, many key steps, such as the selection of an appropriate descriptor subspace, the choice of clustering algorithm, the tuning of parameters, and the validation of the resulting clusters, require human’s vision-reasoning ability, namely, making decisions based on visual inspection and chemical knowledge. Consequently, even advanced automated pipelines ,,, still highly rely on continuous human intervention for visual validation and parameter tuning. In short, human experts remain essential because of their visual perception and domain-specific knowledge. As a result, such analysis remains accessible only to specialists.

This difficulty is not unique to NAMD analysis. Many areas of chemistry, interpreting spectroscopic fingerprints, segmenting microscopy images, monitoring reaction progress from video, or analyzing ground-state molecular dynamics, face the same need to combine visual observation with domain-specific reasoning. In all these cases, the human expert performs an iterative loop of “see, judge, act, and re-evaluate” that has proven difficult to formalize into conventional purely computational algorithms. As a result, the full analysis workflow remains largely inaccessible to nonspecialists and does not scale to the growing volume of simulation data.

Recent advances in large language models (LLMs) opened new possibilities for scientific workflow automation. − However, conventional LLM-based approaches remain poorly aligned with the demands of NAMD analysis. , Most existing applications treat LLMs either as autonomous planners or as passive executors; neither suits trajectory interpretation, which demands simultaneous visual comprehension and chemical reasoning about complex spatial patterns. Text-only LLMs cannot access the visual cues essential for understanding molecular geometries and reaction pathways, leaving this domain largely unexplored.

To address these challenges, we present VisU, a framework driven by Vision-Language Models (VLMs) and designed to automate the unsupervised analysis of NAMD trajectories. VisU adopts a “Mentor–Engineer–Student” paradigm that mimics the collaborative workflow of a theoretical chemistry research group. This framework integrates two recently released state-of-the-art large language models, i.e. Doubao-Seed-1.6-Vision (released September 30, 2025) and DeepSeek-V3.2 (released November 1, 2025), to provide a synergistic approach to data analysis. Within this framework, a Mentor provides physical intuition through visual reasoning and domain knowledge. An Engineer handles adaptive tool-calling and workflow coordination, while a Student executes computational tasks, generates visualizations, and submits results for evaluation. This design establishes a research paradigm that integrates multimodal reasoning, visual inspection, and chemical expertise within a fully automated, four-stage workflow for analyzing the NAMD simulation data.

By embedding vision-capable LLMs into this structured framework, VisU implements an end-to-end pipeline that bridges raw numerical data and chemical knowledge. In this framework, visual reasoning is not merely a supplementary tool but the primary driver of the iterative process, mimicking how a scientist’s eye discerns complex patterns to refine hypotheses and validate mechanisms. This work demonstrates that embedding autonomous visual intelligence into NAMD analysis offers a promising approach for processing excited-state dynamics. VisU combines a fully automated end-to-end workflow, a role-based multiagent architecture, and vision-driven decision-making. It reduces days of manual effort to minutes and makes the analysis accessible to nonspecialists and researchers in related fields.

More broadly, VisU provides a generalizable methodological framework for human-AI collaboration that extends beyond theoretical chemistry. By formalizing the tight coupling of perception, reasoning, and domain expertise, this protocol facilitates more accessible, reproducible, and efficient analysis of complex chemical data. This framework is designed not to replace human insight, but to automate the iterative visual-reasoning loop that is commonly required in data-intensive molecular sciences.

Method and Protocol

In this section, we provide an overview of VisU, an autonomous, vision-AI-powered framework that serves as a comprehensive protocol for unsupervised analysis of NAMD trajectories. The following sections detail the operational procedure and the design principles necessary for researchers to implement and adapt this workflow for their own studies. After carefully studying the organization style of a theoretical group, we propose the whole working protocol of VisU.

In a traditional theoretical team, three roles perform distinct but complementary functions in a joint project: a mentor supervises the overall analysis and provides expert chemical guidance, a programming-engineer constructs computational scripts and integrates them, and a student executes computational tasks while generating intermediate results.

The VisU framework adopts a “Mentor–Engineer–Student” architecture, in which these three roles are designed to mirror their correspondences in a real nonadiabatic dynamics research group.

• The Mentor : Powered by the Doubao-Seed-1.6-Vision LLM, the Mentor integrates deep chemical insights with advanced vision-reasoning capabilities to oversee the analytical direction. By “vision-reasoning”, we mean the ability to read images directly and make decisions based on what is seen, without converting images into data or stopping for human input.

• The Engineer : Built upon the DeepSeek-V3.2 LLM equipped with specialized computational tools and coding templates, the Engineer demonstrates robust programming and debugging proficiency.

• The Student : Responsible for data management, script execution, and visualization, the Student serves as the operational hub, providing feedback to the Mentor and Engineer.

Following this human-mimetic paradigm, VisU operationalizes NAMD analysis into a fully automated, four-stage end-to-end protocol:

• Preprocessing. This stage requires the collaboration of the Engineer and the Student. The Engineer is responsible for the script development and code debugging, while the Student primarily executes them to extract relevant structures from simulation outputs.

• Recursive channel discovery. This phase focuses on identifying nonadiabatic channels through recursive unsupervised ML analysis. The Student operates the analysis pipeline, generates plots, and submits the results for the Mentor’s review automatically. The Mentor then evaluates these plots using visual reasoning to identify the underlying nonadiabatic events, subsequently providing precise guidance on ML algorithm selection and hyperparameter tuning.

• Analysis of dominant molecular motions. Once the channels are identified, the Mentor and Student work together to clarify the dominant molecular motions for each decay pathway. The Student performs specialized programs for downstream analysis. The Mentor interprets the key molecular motions by examining the dimensionality reduction results provided by the Student, and understands the mechanisms of these findings based on chemical insights.

• Final validation and mechanistic summarization. At the end, VisU delivers a highly professional academic presentation through the joint work between the Student and the Mentor. The Student organizes and submits all comprehensive findings, including publication-quality visualizations, to the Mentor. The Mentor conducts the final validation of mechanistic consistency and synthesizes the submitted materials into the final structured report and interpretive representation.

VisU requires only minimal input from the user: a compressed archive containing the raw NAMD simulation outputs (or a few output files with geometric and hopping information) together with a brief natural-language hint describing the file contents. Based on this input, VisU automatically runs the entire analysis pipeline without human intervention, which includes data preprocessing, recursive channel discovery, identification of dominant molecular motions, and final validation. At the end of the workflow, VisU delivers a comprehensive, publication-ready academic report that summarizes the identified nonadiabatic channels, their characteristic nuclear motions, and the associated mechanistic insights. A detailed description of each stage is provided in the following subsections.

By integrating multimodal reasoning, visual understanding, and domain-specific knowledge, VisU mirrors the heuristic intelligence of a research team. It establishes a fully automated end-to-end protocol that uncovers nonadiabatic channels and their underlying mechanisms without manual intervention. The overall framework is illustrated in Scheme , and each component is described in detail in the following subsections.

1. Workflow of VisU, an Autonomous Vision-AI Framework for Hierarchical Analysis of Nonadiabatic Dynamics .

^a The VisU system comprises the three-role paradigm and four functional modules: (a) “Mentor–Engineer–Student” architecture; (b) Preprocessing module autonomously parses raw NAMD outputs and generates standardized descriptor sets; (c) Recursive Channel Discovery module iteratively applies PCA-guided clustering to uncover chemically distinct decay channels; (d) Important-Motion Identification module analyzes PCA projections and visualizes dominant nuclear displacements for each channel; (e) Validation and Mechanistic Summary module selects representative structures, performs final quality checks, and generates concise mechanistic interpretations.

This framework is highly versatile and transferable. The three roles (Mentor, Engineer, Student) work together seamlessly, with each step executed automatically according to the task requirements. This includes automatic invocation of LLMs, generation and execution of analysis scripts, and transfer of data and intermediate results between roles, all enabled by carefully engineered prompts and a library of reusable code templates.

These prompts embed task-specific heuristics derived from computational chemistry practice. They guide the Mentor to perform a series of chemical reasoning steps: identifying ring atoms from a molecular plot, verifying that the observed structural features match the corresponding internal coordinates, and evaluating whether clustering results are chemically reasonable. In addition, the prompts encode domain knowledge of nonadiabatic dynamics, such as recognizing hopping patterns and validating whether separated clusters correspond to distinct reaction channels. They also incorporate general data analysis strategies, including selecting appropriate descriptor subspaces, choosing clustering algorithms, and tuning parameters. The prompts are designed to be general and task-agnostic, rather than optimized for a specific molecular system or reaction. The complete set of prompts is provided in the Supporting Information (SI) and is also available at our GitHub repository.

To ensure broad compatibility, VisU can read and process raw outputs from different NAMD packages without any modification. The framework is also model-agnostic. The underlying LLMs can be replaced or upgraded, such as using different vision-language models for the Mentor or different code-generating models for the Engineer. Finally, VisU is applicable to different types of nonadiabatic dynamics problems, such as dissociation, isomerization, and large-scale molecular structural changes.

The implementation details of each stage are described in the following subsections. Further discussions and additional examples are provided in the SI.

Data Preprocessing

In the first step of VisU, the Data Preprocessing module performs two main tasks. It generates geometry and other relevant files in a standardized format for subsequent analysis, and converts the geometric information into molecular descriptors. This stage involves close coordination among the Mentor, Engineer, and Student, combining chemical insight, reliable programming ability, and automated execution.

Input Files Processing

VisU accepts two input formats: (i) three mandatory XYZ files, that is, ref.xyz (reference geometry of ground-state minimum), hopping.xyz (all hopping geometries from NAMD trajectories), and init.xyz (corresponding initial geometries), or (ii) a single ZIP archive containing raw simulation outputs from any NAMD package. In addition to the files, the user provide a short natural-language hint that describes the contents of the files and the hopping information (e.g., which state change defines a hopping event). Details of such hints and examples are given in the SI.

When a ZIP archive is provided, the Data Preprocessing module first performs trajectory validation. This includes checks for statistical convergence and energy conservation, which are visualized through plots automatically generated by the Engineer. The results are then visually inspected by the Mentor to ensure reliability. This validation step is critical because raw simulation outputs may contain incomplete or corrupted trajectories; the automatic error detection and correction mechanism ensures that only reliable data proceed to subsequent analysis. Only after passing these validations are the structural geometries extracted and organized into the three standardized files.

In principle, one could write custom scripts to handle each NAMD package individually. However, the output formats and directory structures differ substantially across distinct packages, making a fully handmade coding solution labor-intensive and difficult to maintain. VisU addresses this challenge by combining large language models with a collaborative Engineer-Student workflow. The Engineer develops preprocessing scripts based on predefined Python templates and validation tools, while the Student executes the scripts and provides iterative feedback. A detailed demonstration of handling heterogeneous outputs is provided in SI Case Study 1.

Structure Analysis and Descriptor Generation

Following preprocessing, VisU proceeds to systematic structure analysis and descriptor generation through the close cooperation of the Mentor (built on Doubou-Seed-1.6-Vision) and Student. The core of this stage is the introduction of a dedicated vision-language chemistry Mentor, a large multimodal model guided by prompts incorporating chemical knowledge and task-specific heuristics. Extensive prompts are embedded throughout the framework to ensure that the LLM calls are general and adaptable, allowing the same workflow to handle different molecular systems without case-specific tuning. The workflow starts with the Student generating a high-quality 2D molecular diagram from the optimized reference geometry in ref.xyz.

After an initial validation of the reference structure, the Mentor analyzes the molecular building blocks and automatically partitions atoms into chemically meaningful groups, such as ring and nonring atoms. With the multimodal understanding of bonding topology and visual patterns, it identifies and enumerates atoms belonging to aromatic, alicyclic, and fused ring systems. This fully automated, expert-level structural identification eliminates manual intervention while ensuring the generation of chemically interpretable descriptors.

The VisU framework is designed to handle a broad range of photoinduced nonadiabatic dynamics problems, including ring puckering, photoisomerization, and photodissociation. This generality demonstrates that the framework is not limited to a specific type of molecular system or photochemical mechanism. While the current implementation was tested on polyatomic molecules (both ring-containing and bond-breaking systems), extensions to more complex systems or additional photophysical processes will be addressed in future work.

As previous works , demonstrated that redundant internal coordinates (RICs) , provide an appropriate molecular descriptor for NAMD analysis, we adopt RICs here to maintain generality across different molecular systems. RICs consist of bond lengths (R), bond angles (A), and dihedral angles (D), which capture different physical aspects of molecular structure. For systems containing aromatic or alicyclic rings, RICs associated with the ring framework and those involving terminal groups play fundamentally different roles during nonadiabatic dynamics. Accordingly, within the workflow, the Mentor instructs the Student to classify atoms into ring (ring) and end-group (eg) subsets. The Student then constructs six descriptor groups, R _ring, A _ring, D _ring, R _eg, A _eg, and D _eg, which are subsequently used for unsupervised analysis. Hydrogen atoms can be included or excluded depending on the user’s choice; by default, they are excluded to avoid overemphasizing large-amplitude motions. For general systems where no ring appears or no clear ring/end-group distinction exists, the corresponding ring descriptors are set to zero to maintain a consistent input format. The user can specify these choices through the natural-language hint, and VisU automatically adjusts the descriptor construction accordingly. More detailed information is given in SI.

Recursive Channel Discovery

The primary objective of this stage is to examine the distribution patterns of hopping geometries and identify the distinct reaction channels within the NAMD trajectories. This is achieved through a combination of dimensionality reduction and clustering analysis, executed via a collaborative Mentor–Student workflow. With the multimodal capabilities of the vision-LLM, the Mentor acts as a chemical data analyst, providing expert guidance to the Student, who serves as the primary executor.

The “PCA-then-clustering” protocol follows an iterative loop:

• Dimensionality reduction. The Mentor instructs the Student to perform PCA on all hopping geometries using six distinct descriptor families and to present the resulting 2D projection figures for validation. Further details on the descriptor definitions are provided in the SI.

• Visual selection. The Mentor visually inspects these figures to evaluate their chemical significance and identify the low-dimensional representation that exhibits the most discriminative grouping.

• Clustering execution. Based on the identified patterns, the Mentor selects an appropriate clustering method and specifies its initial parameters. The Student then applies the chosen algorithm to partition the hopping geometries.

This three-step protocol is applied recursively. After each clustering event, the Student extracts each resulting subgroup as a new data set and initiates the protocol again. This cycle continues recursively until a cluster is deemed nonseparable or its size falls below a predefined threshold (typically 15% of the total hopping geometries). Upon completion, each terminal cluster is identified as a single candidate nonadiabatic channel. In implementation, this recursive procedure creates a hierarchical tree that enables the systematic enumeration of all reaction pathways, significantly enhancing the modularity and robustness of the automated analysis framework.

The efficiency of this protocol relies on three critical decision-making tasks: (i) selecting the optimal reduced descriptor set; (ii) choosing the most suitable clustering algorithm; and (iii) fine-tuning clustering parameters. Although these tasks are largely intuitive for experienced researchers, they are extremely difficult to automate using rigid mathematical rules. By incorporating vision-enabled reasoning, the Mentor can directly inspect visual representations of the data and identify salient patterns that would otherwise be inaccessible to purely numerical criteria. This capability allows the Mentor to evaluate projections and clustering outcomes in a manner closely resembling the visual assessments routinely performed by human theoretical chemists. Consequently, subsequent analytical decisions are guided by visually grounded scientific judgment rather than fixed heuristics.

The selection of an appropriate descriptor subset follows a standardized visual screening procedure. The Student generates six low-dimensional projections corresponding to all descriptor groups and submits them to the Mentor. Based on the separability of the data distributions, the Mentor identifies the reduced spaces that exhibit the most distinguishable channel patterns and selects the corresponding reduced space for subsequent clustering analysis.

To facilitate the selection of the clustering algorithm, the Mentor is provided with a set of expert-defined principles via specialized prompts. By “perceiving” the data distribution patterns and evaluating them against these principles, the Mentor determines the optimal strategy among three supported methods: K-Means, − Density-Based Spatial Clustering of Applications with Noise (DBSCAN), − and agglomerative clustering. , Further details of these algorithms can be found in the SI.

A particularly critical challenge lies in the dynamic adjustment of clustering parameters. While humans can instinctively evaluate clustering quality through visual inspection, formalizing this intuition into a unified mathematical objective function remains nontrivial. VisU overcomes this obstacle by including vision-reasoning to address two distinct operational scenarios.

For algorithms such as K-Means or agglomerative clustering, parameter selection (e.g., the number of clusters, k) can be derived from the visual distribution of the data. In these cases, the Student performs an initial execution using randomized or default parameters to produce a preliminary labeled scatter plot. The Mentor then inspects this visualization and directly dictates the necessary adjustments to align the results with the perceived data distribution.

For algorithms like DBSCAN, a more complex challenge arises: its core parameters (eps and min_samples) lack a direct visual interpretation. Even for human experts, determining optimal values through a single visual inspection is nearly impossible. To address this, VisU implements a reasoning-driven strategy that combines expert guidance with iterative evaluation. The Mentor is first informed of the meaning of each parameter and general adjustment principles. Then, by observing the sensitivity of clustering outcomes to parameter variations, the Mentor and Student engage in a multiturn interaction, systematically refining parameter values based on visual feedback and logical reasoning until a satisfactory partition of reaction channels is obtained.

Important-Motion Identification

Following the identification of all involved nonadiabatic reaction channels, the next task is to elucidate the active molecular motions that drive the system from the Franck–Condon region to the conical intersection. PCA is used for this purpose. The analysis employs the same six descriptor groups defined in the preceding Channel Discovery stage. For each channel, the hopping geometries and their corresponding initial structures are combined. By examining whether initial and hopping geometries exhibit clear spatial separation in the PCA projections, the descriptor sets primarily responsible for the nonadiabatic decay can be clarified. Subsequent analysis of the contribution of individual descriptor elements to the principal components allows for the identification of the active coordinates governing the dynamics. This process relies on the interpretation of PCA results, a task that is intuitive for human researchers but difficult to formalize with conventional computational routines.

The identification of these critical motions is performed through a tight coordination between the Mentor and the Student. For each identified channel, the Student assembles the initial and hopping geometries, executes PCA, and generates a comprehensive package. This package includes 2D projection plots, heatmaps of the principal components, and three-view (front, side, and top) renderings of both the reference and averaged hopping structures. Under the guidance of its vision-reasoning capabilities, the Mentor evaluates these figures to identify the leading molecular motions. This Mentor-Student joint procedure effectively bridges the gap between unsupervised data clustering and human-interpretable chemical mechanisms, establishing a fully automated NAMD analysis protocol that eliminates the requirement for manual validation.

Validation and Mechanistic Summary

In the final stage, VisU consolidates all analytical findings into a comprehensive analysis report. This comprehensive synthesis is achieved through the integration of the Mentor’s visual reasoning and chemical expertise, the Engineer’s specialized programming capabilities, and the Student’s disciplined execution.

For each reaction channel, the Student randomly samples a representative set of hopping geometries (typically ten) and prepares their three-view (front, side, and top) renderings for the Mentor. Upon visual evaluation, the Mentor identifies the most characteristic geometry based on previously established analysis results. Simultaneously, the Mentor inspects representative geometries across all channels to identify and merge mirror-symmetric pairs, ensuring a concise and physically sound classification. Following the selection of representative geometries, the Engineer and Student collaborate to generate trajectory evolution animations. The Engineer adapts and refines predefined code templates to construct robust visualization scripts, which are then executed by the Student to produce high-fidelity videos for the visualizations of typical trajectory evolution.

Finally, all generated assets, including analysis recordings, annotated figures, and optional video animations, are forwarded to the Mentor. At this juncture, the Mentor leverages both its visual perception and deep chemical insights to perform a systematic evaluation. Through rigorous domain-specific prompt design, the Mentor is equipped with the specialized knowledge of a senior computational chemist. This allows the Mentor to re-examine the entire workflow: identifying reaction channels, tracing dominant atomic motions, correlating them with active coordinates, and performing cross-channel comparisons.

The multimodal assessment capability of the Mentor facilitates the generation of a concise, human-readable summary for each channel. These summaries, paired with key visualizations, are compiled into a mechanistic report. This final stage elevates VisU from a mere unsupervised analysis engine to an autonomous computational collaborator that not only extracts essential features from NAMD simulations but also delivers expert-level chemical insights.

Overall, by integrating vision-capable LLMs within a systematic prompt framework, VisU performs an end-to-end analysis of NAMD simulation data and produces finalized scientific reports with complete autonomy. This demonstration proves that vision-driven multimodal intelligence provides an effective solution for complex scientific problems that traditionally require both human visual intuition and domain-specific expertise.

Case Study

To illustrate the practical utility of the VisU workflow, we applied the framework to representative NAMD systems and evaluated its performance in analyzing their molecular motions and reaction pathways. The results confirm that this protocol is straightforward to operate and broadly applicable across diverse NAMD problems.

The VisU framework is applicable to a wide range of nonadiabatic dynamics problems, including photodissociation, ring puckering, and isomerization. In the following, we use keto isocytosine as a prototypical system to illustrate the analysis workflow and its outcomes in detail, focusing on the photoinduced reaction channels and ring deformation motions. This example demonstrates the VisU provides the reasonable analysis reports with high efficiency. In addition, we tested VisU on another system, CH₂NH₂ , which involves photodissociation and torsion. The successful applications of VisU in these examples demonstrate the generality of the framework. Due to space limitations, we focus on keto isocytosine in detail and present the analysis of CH₂NH₂ in the SI.

As a tautomer of cytosine, one of the canonical DNA bases, the photoinduced dynamics of keto isocytosine was extensively investigated both experimentally and theoretically. − The molecular structure of keto isocytosine with atom labeling is shown in Figure .(d). The NAMD simulation data used in this study were taken from our previous work, and relevant details are given in SI as well.

1.

Preprocessing pipeline implemented through the cooperation of Mentor, Engineer, and Student. (a) User-provided input file (test.zip) and message. (b) Generation of the preprocessing Python script using tool calling and the provided Python function templates. (c) Final preprocess.py and the resulting XYZ files. (d) Analysis of molecular structures, including the molecular structure and atomic labels of keto isocytosine shown in mol.png. (e) Six generated descriptor sets based on RICs.

Preprocess the Input File

Preprocess the Input ZIP File

To demonstrate the end-to-end autonomy of VisU, we deliberately chose the raw, unprocessed NAMD simulation output from the JADE, ,, package. At the end of the pipeline, VisU produces a complete analysis report that includes identified reaction channels, dominant molecular motions, and mechanistic interpretations. JADE is used here only as an example; the same workflow applies to outputs from other NAMD packages without modification.

This step requires the raw NAMD simulation outputs and a short natural-language hint from the user. We demonstrate the workflow using the JADE package and the keto isocytosine system. The user provides a compressed archive containing the JADE output files and a brief hint (Figure (a)). No manual preprocessing of the input files is needed. The hint is interpreted by the large language model behind the Engineer, which then automatically executes the preprocessing workflow through carefully designed prompts and a library of reusable code templates (Figure (b)). The operations include unzipping files, validating contents, collecting initial geometries, identifying hopping times, and extracting the corresponding hopping geometries. Although JADE is used here as an example, the same pipeline can read raw outputs from other NAMD packages without modification, demonstrating its general applicability. A detailed discussion is provided in SI Case Study 1.

With only minimal input hints, VisU demonstrates strong data-preprocessing capability by automatically handling a previously unseen NAMD output archive. This successful application can be attributed to two essential features of VisU: the tool-aware reasoning capability enabled by DeepSeek-V3.2 and the well-defined reference templates provided by the developers.

Before extracting the structural information, VisU performed the initial check of the statistical convergence of the trajectories to avoid potential errors. The Engineer generated Python scripts to perform bootstrap resampling and produce population plots. The Student executed these scripts, obtaining population layout figures, which were then inspected by the Mentor to confirm convergence.

As discussed in the previous section, this preprocessing step produced three required output files, namely ref.xyz, init.xyz, and hopping.xyz (Figure .(c)).

Generate the Descriptor Set

To generate molecular descriptors, VisU processed both the reference (ref.xyz) and hopping (hopping.xyz) geometries. The Student plotted a 2D molecular image (mol.png) with atomic labels on the basis of the reference geometry, as illustrated in Figure .(d). This image was then passed to the Mentor, powered by the Doubao-Seed-1.6-Vision model, which was carefully prompted to understand the context of the analysis workflow.

With the visual-reasoning capability, the Mentor first interpreted the image to determine whether the input molecule was compatible with the current available protocol. If not, the error messages are provided and the whole workflow stops. For the present system, VisU identified the atoms constituting the ring moiety, which provided explicit guidance for the Student to construct the corresponding RIC descriptors. Using the hopping geometries collected in hopping.xyz, the Student then built six sets of RIC descriptors, as shown in Figure .(e): D _ring, A _ring, R _ring, D _eg, A _eg, and R _eg. These descriptors formed the foundation for subsequent unsupervised channel discovery and mechanistic analysis. In parallel, the same descriptor sets were computed for all initial geometries in init.xyz.

Extract the Channels

Using the prepared descriptor sets, channel discovery proceeded via an iterative PCA-then-clustering analysis protocol.

As shown in Figure .(a), the first iteration of channel discovery began by applying PCA to each of the six descriptor sets to obtain low-dimensional representations of the hopping geometries. In all cases, the first two principal components contributed more than 60% of the total variance. The resulting PCA projections were then forwarded to the Mentor for assessment. In the present example, the Mentor leveraged its visual-reasoning capability to identify that the hopping points were well separated in the D _ring-based reduced space, indicating the presence of several distinct channels and highlighting the important role of ring-associated dihedral motions in the nonadiabatic dynamics.

2.

“Mentor–Student” cooperation during the first iteration of unsupervised channel discovery. Based on diagnostic plots produced by the Student, the carefully prompted visual-language Mentor (Doubao-Seed-1.6-Vision) provides targeted guidance on: (a) selection of the optimal reduced-dimensional subspace, (b) choice of clustering algorithm, (c) tuning of hyperparameters, and (d) final evaluation of the resulting clusters.

The next critical step is to group similar hopping geometries into clusters. In Figure (b), the Mentor inspected the distribution in the D _ring PCA plot using its visual-reasoning capability and recommended agglomerative clustering for the initial partitioning. As shown in Figure (c), the Student first performed clustering with the default parameter k = 2. Upon reviewing the resulting plot, the Mentor immediately identified that the central region was visually distinct and was not merged with the adjacent cluster. Therefore, the Mentor instructed the Student to adjust the parameter to k = 3, after which the clustering result aligned with expectations. Following reclustering with the corrected parameter, the Mentor confirmed that the current clustering was accurate.

A final visual inspection by Mentor, shown in Figure (d), verified that the current clustering result with three clusters was reasonable, ensuring that this step aligned with expert-level visual intuition. The first discovery iteration thus concluded with three clusters: Cluster 0, Cluster 1, and Cluster 2. Among them, Cluster 2 contained only 37 structures, corresponding to less than 15% of the total data set. This cluster was therefore treated as an independent candidate channel (Candidate Channel 0), while the remaining two clusters proceeded to a second iteration of analysis. Afterward, the same PCA-then-clustering procedure was recursively applied to all remaining subsets. As illustrated in Figure S1, all clusters either became nonseparable after three iterations or contained fewer than 15% of the structures, ultimately yielding seven nonadiabatic decay channels (Candidate Channel 0 to 6).

As different data patterns appeared in these iterative steps, distinct clustering algorithms were recommended based on their distributions: including agglomerative clustering (Iteration 1), DBSCAN (Iteration 2), and K-Means (Iteration 3), as in Figure S1. Further results and the associated rules used in the prompts can be found in the SI. These observations demonstrate that VisU can effectively handle both algorithms like agglomerative clustering, which requires only a single hyperparameter and aligns closely with visual intuition, and more complex methods such as DBSCAN, where the two parameters (eps and min_samples) are less directly interpretable from visual inspection.

Identify the Important Molecular Motion

After extracting the candidate channels, the important molecular motions associated with each channel were identified. Candidate Channel 2 is presented here as a representative example to illustrate the analysis workflow and the resulting mechanistic interpretation (Figure ).

3.

Identification of the dominant molecular motions, illustrated for Candidate Channel 3. (a) Two-dimensional PCA projections generated by the Student from the structural descriptors D _ring, A _ring, R _ring, D _eg, R _eg, and A _eg. By visually inspecting these projections, the Mentor assessed in which low-dimensional representations the initial and hopping geometries became separable. (b) Corresponding PCA component loadings, examined by the Mentor to identify the structural features most responsible for the observed channel separation. (c) Dominant molecular motions inferred by the Mentor through joint inspection of the key features and representative initial and hopping geometries.

At this stage, the Student performed PCA projections of both the initial and hopping geometries, submitting the resulting two-dimensional representations to the Mentor for analysis. By visually inspecting these projection plots, the Mentor determined which sets of RICs were most effective for distinguishing hopping from initial geometries. In the present example (Figure (a)), the hopping and initial geometries were clearly separated in the reduced space defined by D _ring, D _eg, A _ring, and R _ring, indicating that these RIC sets captured the leading reactive coordinates in the photochemical processes.

Subsequently, the Student computed the contribution of each RIC to the principal components and visualized the results in Figure (b). Upon reviewing these plots, the Mentor identified the most significant reactive coordinates, including D­(4,1,5,3), D­(10,4,1,5), and D­(4,10,2,3) in D _ring; D­(6,1,5,3) and D­(10,4,1,6) in D _eg; A­(1,4,10), A­(1,5,3), and A­(4,1,5) in A _ring; as well as R­(1,5), R­(3,5), and R­(2,3) in R _ring.

As shown in Figure .(c), using these key features, the Mentor further identified that the dominant molecular motions governing Candidate Channel 2 are C1 puckering and O9–C10 out-of-plane deformation. With these identified critical motions, the Mentor then selected a representative geometry for mechanistic inspection from several randomly chosen hopping snapshots collected in the previous steps.

After the Mentor selected representative hopping geometries, the Engineer generated a Python script to extract the corresponding trajectory structures and produce videos for visual inspection of the molecular evolution in each channel.

The same analysis procedure was systematically applied to all candidate channels, enabling the identification of their respective dominant molecular motions, representative structures, and corresponding trajectory animations. All additional details are provided in the SI.

Validate and Summarize Mechanisms

Up to this point, seven candidate nonadiabatic channels and their characteristic molecular motions were identified. The next step involves the Mentor evaluating the results to assess the rationality of all discoveries and to prepare the final analysis report.

An essential task for the Mentor is to check for chirality. For example, several candidate channels in the present system were identified as mirror counterparts based on both qualitative visual inspection and quantitative data comparison. As shown in Figure (a), the Mentor merged different groups showing mirror-symmetry: Candidate Channel 1 and Candidate Channel 6 into Channel B; Candidate Channel 2 and Candidate Channel 4 into Channel C; and Candidate Channel 3 and Candidate Channel 5 into Channel D. Candidate Channel 0 exhibited no mirror-symmetric counterpart and was therefore retained as Channel A. As in Figures (b) and (c), the initial seven candidate channels were finally consolidated into four distinct nonadiabatic channels: Channel A, B, C, and D.

4.

(a) Final nonadiabatic channel analysis tree. Details regarding the division and naming of different clusters can be found in the SI. (b) Averaged and (c) representative structures (hydrogen atoms omitted) for the four channels.

It is important to note that different decay channels may be governed by the same nuclear motions but with varying magnitudes. For instance, Channel B, C, and D all exhibited pronounced out-of-plane motions of O9 and C10, while their directions differed. In principle, one can perform systematic pairwise comparisons of initial and hopping geometries across channels; however, this approach becomes tedious in multichannel situations. Instead, VisU relies on the Mentor to perform a holistic comparison by integrating key features, representative geometries, and dominant motions across all channels. Through direct visual inspection and cross-channel assessment, the Mentor clarifies the mechanistic similarities and differences among the nonadiabatic decay channels.

Finally, the Mentor drew the following conclusions. Channel A accounted for approximately 9% of the total hopping events, which was characterized by a dominant C9–C10 stretching motion. In contrast, the remaining three channels involved combinations of C1 puckering and O9–C10 out-of-plane motions. Among them, Channel B, Channel C, and Channel D contributed approximately 10%, 75%, and 6% of the total population, respectively. Although these three channels shared similar core motions, they differed in the directionality of the side-group motion associated with the O9–C10 fragment, as shown in Figures .(b) and (c). These results are consistent with our previous and recent studies, , and additional discussions are provided in the SI.

At the end, the complete mechanistic interpretation, including channel populations, representative structures, and visual summaries derived from the raw nonadiabatic dynamics trajectories, was compiled in a comprehensive analysis report provided in the SI. With this, the end-to-end analysis performed by VisU is complete, which was executed fully automatically without human intervention.

The above demonstration uses keto isocytosine as a specific example. To show that VisU is general, we also tested it on several additional cases. First, we applied VisU to CH₂NH₂ , a system that contains no ring structure. Its nonadiabatic dynamics involves both photodissociation and double-bond torsion. VisU successfully identified the key nuclear coordinates governing its excited-state decay. Second, VisU can read raw outputs from multiple NAMD packages, including JADE, ,, MNDO, Newton-X, and MLatom, , without any workflow modification. This flexibility is achieved by combining LLMs, user-provided hints, and a library of reusable code templates. Third, the framework is model-agnostic, as it supports different underlying LLMs (e.g., alternative vision-language models or code-generating models). All relevant tests are described in the SI (Case Studies 1 and 2). These results confirm the general applicability of VisU to different molecular systems, dynamical processes, software environments, and LLM backends.

Practical Considerations

With the application of the VisU protocol demonstrated through the case study above, we now turn to several important considerations for its broader use. The practical utility of any methodological protocol depends not only on its core functionality, but also on its efficiency, applicability, robustness, limitations, and practical deployment. In this section, we discuss these aspects in details.

VisU provides an integrated and efficient workflow for nonadiabatic dynamics analysis. Traditionally, a purely manual analysis requires the researcher to test many possible internal coordinates, examine numerous plots, and repeatedly adjust the analysis based on trial and error, making the workflow stop-and-go and typically taking several days. Semiautomated workflows using dimensionality reduction and clustering reduce some manual work but still need human intervention for descriptor selection, parameter tuning, and result validation, often still requiring days of researcher time. VisU is not designed to replace human experts or to be more accurate than standard methods. Instead, it automates the entire process from raw simulation outputs to final report generation with minimal user input. In the keto isocytosine case, the complete analysis takes approximately 15 to 20 min on a standard desktop computer, reducing total analysis time from days to minutes by integrating visual perception and decision-making into a seamless loop. Moreover, VisU can process multiple independent trajectory sets simultaneously, producing complete reports without additional human effort, a capability impossible in manual or semiautomated workflows. This automation makes the analysis faster, more reproducible, and accessible to nonspecialists, and its results agree with previous studies. For the keto isocytosine example, this work builds on our previous semimanual protocol and reproduces its conclusions; see the SI for a detailed comparison. Visual reasoning here plays an organizing role to guide the analysis process rather than providing new physical insight.

A key structural innovation of VisU is its Mentor–Engineer–Student multiagent architecture, which mimics the division of labor in a human research team. The Mentor, a vision-language model, provides expert chemical reasoning. The Engineer, a code-generating LLM, constructs and debugs scripts. The Student executes tasks and manages data. This role-based collaboration enables the framework to handle complex, multistage analyses that would be difficult to realize with a single monolithic LLM. The reasoning of the Mentor is guided by prompts that encode task-specific heuristics rather than raw experimental data or extended theoretical discussions. These heuristics cover common analysis decisions, such as checking trajectory convergence, validating intermediate results, selecting representative structures, and merging mirror-symmetric channels. The complete set of prompts and heuristics is provided in the SI.

A key feature of VisU is the integration of visual perception and decision-making. The Mentor reads images and videos directly, interprets PCA plots, selects clustering parameters, executes the clustering, and evaluates the results, all without human intervention. Importantly, this visual reasoning operates directly on raw, unprocessed images, without the need for complex preprocessing pipelines. Visual reasoning is not a mere supplement but the primary driver of the iterative process, directly connecting what the model observes to the decisions it makes. This tight coupling enables a seamless analysis flow and eliminates the stop-and-go nature of manual or semiautomated workflows.

The framework also incorporates dedicated error handling and reliability checks. Input data are automatically examined for convergence, energy conservation, and file integrity. Simple format errors can be corrected automatically. Critical decisions are cross-checked using prevalidated code templates. If script execution fails, the traceback information is fed back to the Engineer for iterative self-repair. If a user suspects an error, the pipeline can be restarted. All automatic decisions, together with the reasoning behind them, are documented in the final report. Further details are provided in the SI.

VisU further demonstrates broad generality. The framework is applicable to diverse molecular systems, including molecules with or without rings and those undergoing bond breaking or isomerization. It accepts outputs from different NAMD packages such as JADE, ,, MNDO, Newton-X, and MLatom. , It handles different types of nonadiabatic dynamics problems, including dissociation, ring puckering, and torsion. Moreover, the framework is model-agnostic: the underlying LLMs can be replaced or upgraded, and we verified that alternative vision-language models and code-generating models work consistently (see SI).

We recognize that PCA is a linear method and may not fully capture highly nonlinear motions. Nevertheless, the use of RICs as molecular descriptors incorporates certain nonlinear information through dihedral angles. The separation of coordinates by type (bonds, angles, dihedrals) and by region (ring, end groups) further helps distinguish different types of motion. For the systems tested in this work, PCA works well. We acknowledge that PCA may not be sufficient for systems with strong nonlinearities. In such cases, nonlinear dimensionality reduction methods could be integrated into the VisU framework without fundamental changes to the workflow. A systematic comparison of different methods is a valuable direction for future work. The main contribution of this study is to demonstrate that visual reasoning can be effectively integrated with unsupervised learning to automate trajectory analysis, and using PCA as an example is sufficient for this purpose.

The vision-reasoning paradigm demonstrated here is not limited to TSH nonadiabatic dynamics. The same workflow can be applied to analyze trajectories from other nonadiabatic methods or even ground-state molecular dynamics simulations. More broadly, similar workflows could be applied to other chemical problems that rely on visual pattern recognition, such as interpreting spectroscopic fingerprints, segmenting microscopy images, or monitoring reaction progress from video.

VisU focuses on nuclear geometric evolution and does not directly incorporate electronic-state properties into its descriptors. Instead, electronic-state information is used as contextual guidance to inform the geometric analysis. Specifically, the analysis is restricted to hopping geometries, i.e., nuclear configurations where electronic transitions occur, typically near conical intersections, which ensures that the structural features examined are directly linked to nonadiabatic events. In addition, at the beginning of the workflow we compute and inspect electronic population plots to verify the reliability of the dynamics. Furthermore, the user can specify which electronic state pair (e.g., S ₁ to S ₀ or S ₂ to S ₁) to analyze through the natural-language hint. These steps connect the geometric analysis to the nonadiabatic events in the simulated trajectories, while the framework does not directly incorporate electronic-state properties into the descriptors.

Nevertheless, several limitations remain. Automatic extraction of detailed electronic character (e.g., orbital shapes or transition types) is not yet feasible in a fully automated pipeline; users can manually add such characterization after the geometric analysis if needed. A brief natural language hint from the user is still required. Performance was mainly validated for gas-phase molecules of moderate size, and more complex systems (e.g., condensed phases or large biomolecules) may still present significant challenges. The recursive PCA-clustering strategy may fail when reduced projections lack clear separability. On the implementation side, the current implementation relies on commercial LLM APIs, so runtime and cost may vary with network conditions and model availability. Detailed discussions of generality, robustness, and current limitations are provided in the SI.

Overall, this case study illustrates the integration of a general-purpose VLM into a domain-specific workflow through task decomposition, information segregation, and coordinated tool use. By combining automation, visual reasoning, and intelligent decision-making, VisU represents a shift from conventional workflow automation toward more robust, scalable, and accessible mechanistic analysis of photochemical dynamics.

Conclusion

We developed VisU, a visual-LLM-driven framework for fully automated, unsupervised analysis of nonadiabatic molecular dynamics (NAMD) trajectories. By emulating the collaborative workflow of a research group through a three-tier “Mentor–Engineer–Student” cooperation, VisU establishes a new research paradigm that systematically integrates visual reasoning with domain-specific chemical knowledge, enabling the identification of dominant decay channels in the nonadiabatic dynamics and their underlying mechanisms of photoinduced reactions.

The powerful analysis capability of VisU stems from the integration of two state-of-the-art large language models: Doubao-Seed-1.6-Vision, which provides expert-level visual reasoning, and DeepSeek-V3.2, which enables adaptive tool-calling and script construction. This synergy facilitates a multiagent paradigm that orchestrates a comprehensive four-module workflow: Preprocessing, Recursive Channel Discovery, Important-Motion Identification, and Validation and Mechanistic Summary.

In a representative case study, VisU successfully identified four distinct nonadiabatic channels, uncovered the corresponding leading reactive coordinates, and automatically handled mirror-symmetry merging. Furthermore, it performed cross-channel comparisons and generated a comprehensive, figure-rich academic report without human intervention. The framework demonstrated a professional-grade ability to reproduce dominant mechanistic features reported in previous work. , The entire analysis, which previously required days of manual effort, now completes in approximately 15 min. Moreover, the framework offers the potential for parallel execution across multiple systems or repeated validation runs. Remarkably, starting only from raw NAMD simulation outputs and minimal text prompts, the entire protocol is executed through autonomous visual inspection and expert-like guidance, powered by advanced vision-reasoning LLMs. These results demonstrate that VisU offers a fully automated, human-like approach for the mechanistic interpretation of NAMD trajectories, substantially lowering the barrier to routine photochemical analysis.

VisU demonstrates that visual reasoning, when embedded within a structured multiagent protocol, can serve as an effective engine for automating complex mechanistic analysis in theoretical chemistry. This framework bridges the gap between qualitative visual patterns, recognized by experts, and quantitative simulation data, by integrating them within a unified, reproducible workflow. By combining multimodal intelligence with domain-specific knowledge, VisU offers a generalizable methodological protocol that goes beyond conventional data processing to emulate aspects of collaborative researchintegrating perception, reasoning, and iterative decision-making for scientific analysis. This vision-driven protocol provides an efficient and transparent approach for analyzing excited-state dynamics simulations and may be extended to other AI-assisted chemical analyses. Ultimately, VisU illustrates the potential of an agentic research paradigm that combines visual reasoning with domain-specific chemical knowledge, offering a new perspective for integrating multimodal intelligence into complex scientific data analysis.

Technical Details

To ensure the reliability, reproducibility, and practical applicability of the proposed methodological framework, we provide the essential technical details for reproducing and adapting the VisU workflow. Further data and Supporting Information are available in the SI and on the accompanying Web site (discussed in the following section).

Framework Implementation and Environment

VisU was implemented in Python 3.12. Core dependencies included Scikit-learn (v1.5.1) for dimensionality reduction and clustering, RDKit (v2025.09.1) for molecular manipulation, as well as Matplotlib (v3.10.7) for plotting.

Molecular geometries in hopping.xyz were aligned to the reference geometry in ref.xyz using the Kabsch algorithm. , Redundant internal coordinates (RICs) were generated using Pysisyphus (v1.0.0) to represent molecular motions efficiently.

All molecular structures were rendered with PyMOL (v3.1.0). Optional trajectory animations were generated as MP4 videos using PyMOL (v3.1.0) and FFmpeg (v7.1.1).

LLM Agents and Workflow Automation

The VisU framework operates in an ”end-to-end multimodal” manner. Here, ”end-to-end” means that all intermediate steps (from raw data preprocessing to final report generation) are handled automatically by the framework. ”Multimodal” means that the framework processes integrate different types of information, including text (user instructions, simulation metadata), images (PCA plots, molecular structures), and code (analysis scripts, visualization commands).

The “Mentor–Engineer–Student” agents were powered by state-of-the-art large language models: Doubao-Seed-1.6-Vision (released September 30, 2025) served as the multimodal Mentor, while DeepSeek-V3.2 (released November 1, 2025) served as the programming Engineer. API calls were handled via the Volcengine Python SDK Ark for the former and the OpenAI Python SDK for the latter.

The choice of these LLM models is driven by functional requirements. The VisU framework requires (i) the ability to directly interpret visual data (e.g., reduced 2D space plots and trajectory videos) without converting images/videos into numerical data first, and (ii) reliable code generation and file-system interaction capabilities. Based on these requirements, we selected Doubao-Seed-1.6-Vision for the Mentor (visual ability) and DeepSeek-V3.2 for the Engineer (script writing and execution). These models offer comparable performance at a lower cost, making them more accessible for routine use.

Doubao-Seed-1.6-Vision provides a substantial advantage in handling complex multimodal reasoning tasks. Leveraging its 256k-token context window, the model could accommodate multiple trajectory projections, long textual descriptions, and optional trajectory videos within a single prompt for comprehensive analysis. Context caching ensured the automatic persistence of historical information, while structured outputs enabled efficient postprocessing and downstream integration. Within VisU, these capabilities allow the Mentor to inspect multiple candidate channels concurrently, evaluate structural patterns, and generate precise, human-like mechanistic guidance.

DeepSeek-V3.2 integrates advanced autonomous tool-calling within a chain-of-thought (CoT) reasoning paradigm. In VisU, it empowers the Engineer agent to adaptively generate and execute context-aware preprocessing scripts and analysis codes, dynamically interact with the file system, and coordinate with Student modules to ensure correct task execution. These features enable the automated handling of heterogeneous simulation outputs, the flexible adaptation to previously unseen data archives, and the efficient construction of all computational codes and scripts.

Within VisU, the embedded prompts follow a consistent design principle, which includes: (i) task description, (ii) input specifications, (iii) relevant domain knowledge, (iv) solution strategies, and (v) output requirements. These prompts are not heavily optimized for specific cases but are intended to apply across various systems and tasks, ensuring that VisU’s performance is driven by its overall workflow rather than specific prompt tuning.

The structural data for the case study, including the hopping geometries and corresponding initial ones extracted from nonadiabatic dynamics, are available on our group’s Web site (https://langroup.site/datasets/). Additional related data and details on the nonadiabatic dynamics simulation settings can also be found on the same Web site. Our source codes are publicly available at the GitHub repository (https://github.com/Yifei-Zhu/VisU.git). To run VisU, users must provide their own API key and service URL by configuring the .env file accordingly. The prompts used in the VisU framework are also available at the GitHub repository.

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/jacsau.6c01054.

• SI-I includes details regarding the methods and related discussions (PDF)

• SI-II presents the results of the case studies, including performance tests on various NAMD software, detailed case study outcomes, and a thorough analysis of CH₂NH₂ (PDF)

• SI-III contains all the prompts used in VisU (PDF)

#.

Yifei Zhu and Jiahui Zhang contributed equally to this work. Yifei Zhu: Conceptualization, Methodology, Software, Formal analysis, Investigation, Data curation, Writing - original draft, Writing - review and editing. Jiahui Zhang: Conceptualization, Methodology, Software, Formal analysis, Investigation, Data curation, Writing - original draft, Writing - review and editing. Chuqiao Feng: Software, Formal analysis, Investigation, Data curation. Binni Huang: Data curation, Investigation, Visualization. Wenyi Yang: Software, Data curation. Zhenggang Lan: Conceptualization, Supervision, Project administration, Resources, Funding acquisition, Writing - review and editing.

The authors declare no competing financial interest.