Artificial Intelligence in Spectroscopy: Advancing Chemistry from Prediction To Generation and Beyond

1 Introduction

The rapid advancements in Artificial Intelligence (AI) and Machine Learning (ML) are reshaping scientific disciplines of chemistry, by streamlining tasks such as molecular property prediction [Guo et al., 2021] and reaction modeling [Coley et al., 2019]. Despite these breakthroughs, the application of ML to spectroscopy-hereafter referred to as Spectroscopy Machine Learning (SpectraML) [Elias et al., 2004; Ralbovsky and Lednev, 2020]-remains relatively underexplored. Spectroscopic and spectrometric techniques, which provide high-sensitivity insights into molecular structure, dynamics, and properties, are now generating large volumes of data due to advances in high-throughput experiments and automated acquisition. Consequently, traditional manual analysis methods, reliant on expert interpretation and reference libraries [Alberts et al., 2024b; Zhu and Jonas, 2023], are increasingly inadequate for handling the scale and complexity of modern spectral datasets.

The growing interest in this field is reflected in the increasing number of research papers that expand the scope of tasks addressed by spectrum-based ML models [Gabriel et al., 2024; Alberts et al., 2024b; Guo et al., 2024; Bushuiev et al., 2024]. However, existing overviews often focus on a single modality (e.g., UV alone or MS) [Beck et al., 2024] or lack a clear framework for distinguishing forward (molecule-tospectrum) from inverse (spectrum-to-molecule) tasks [Sridharan et al., 2022a]. By contrast, our survey unifies five major spectroscopic techniques-MS, NMR, IR, Raman, and UVVis-within a single methodological framework. Moreover, we highlight the rapid progression of spectroscopic analysis driven by ML advances in generative modeling, few- or zeroshot learning, and large-scale pretraining, and we provide an open-source repository of datasets and code. By bridging computational and experimental viewpoints, our work fills a key gap in the literature and highlights new avenues for interdisciplinary collaboration in SpectraML.

The rapid advancements in ML and AI have been transforming workflow automation in spectral analysis, as illustrated by the timeline in Fig. 1. Deep learning models, such as convolutional neural networks (CNNs) [O'Shea, 2015]

and recurrent neural networks (RNNs) [Schuster and Paliwal, 1997], have proven effective in tasks like peak detection, deconvolution, and reaction monitoring. Additionally, transfer learning and pre-trained models [Guo et al., 2024] enable these algorithms to generalize across diverse spectra, thereby reducing the need for extensive retraining. Emerging foundation models [Bommasani et al., 2021] further extend the capabilities of SpectraML by offering advanced reasoning and planning for complex tasks such as molecular structure elucidation and reaction pathway prediction [Guo et al., 2024]. As AI techniques continue to evolve, there is a critical need for a structured discussion on positioning the different capabilities of AI models across various spectroscopy tasks, as well as underscoring key challenges, limitations, and future directions.

This survey addresses these needs with the following contributions:

1. We offer a comprehensive overview of current SpectraML techniques across five major spectroscopic modalities-MS, NMR, IR, Raman, and UVVis-highlighting both methodological innovations and practical applications. Unlike existing surveys that focus on a single modality or overlook the distinction between forward (molecule-to-spectrum) and inverse (spectrumto-molecule) tasks [Beck et al., 2024; Sridharan et al., 2022a], our work provides a unified perspective and frames these tasks within AI's problem-solving role.
2. We present a unified roadmap that traces the evolution of ML in spectroscopy, from early pattern recognition and predictive analytics to advanced generative and reasoning frameworks, thus situating current progress within a broader historical context. It helps researchers understand how foundational techniques have shaped modern approaches and guides them in innovating future methodologies in SpectraML.
3. We identify key challenges (e.g., data quality, multimodal integration, and computational scalability) and emerging opportunities (e.g., foundation models, synthetic data generation, few- or zero-shot learning, and large-scale pretraining) in SpectraML. To facilitate further research, we provide, and will maintain, an opensource GitHub repository containing datasets and code. This work thus serves as a valuable resource for researchers and practitioners in this interdisciplinary field.

The remainder of this paper is organized as follows. Section 2 introduces spectral data representations and the definition of fundamental Forward and Inverse Problems in spectral analysis. Section 3 categorizes and summarize SpectraML approaches in solving forward and inverse problems. In Section 4, we discuss major challenges and highlight emerging directions such as foundation models, and synthetic data generation. Section 5 concludes the work.

2 Background

2.1 Applications of Spectroscopy in Chemistry

Spectroscopy, the study of the interaction between matter and electromagnetic radiation, produces data that resembles audio signals in its representation-peaks, shifts, and patterns that encode molecular information [Elias et al., 2004]. Spectrometry, on the other hand, focuses on measuring chemical interaction to gain insight into molecular structures and properties [Ralbovsky and Lednev, 2020]. Common spectroscopic techniques include mass spectrometry (MS), infrared (IR), Raman, ultraviolet/visible (UV-Vis), and nuclear magnetic resonance (NMR). Each of these techniques is akin to a "lens" providing a different perspective of the molecular world, and when combined, they reveal a fuller picture of molecular structures.

• Mass Spectrometry (MS) allows for the determination of the molecular mass and formula of a compound, as well as some of its structural features by identifying the fragments produced when the molecule breaks apart.
• Infrared (IR) and Raman spectra data allow the identification of the types of functional groups in a compound.
• UV-Vis spectra data provides information about compounds that have conjugated double bonds.
• NMR spectra data provide information about atomic nuclei (e.g., the carbon-hydrogen framework of a compound). Advanced techniques, 2D and 3D-NMR, further enable the characterization of complex molecules such as natural products, proteins, and nucleic acids.

The obtained spectra data are widely used across chemistry, biology, and related fields, akin to a "molecular microscope" that enables researchers to explore the unseen. These spectral data are often presented as plots or graphs that visually represent the relationship between intensity and a specific variable, such as wavelength, wavenumber, or mass-to-charge ratio (m/z), as demonstrated on the left part of Figure 2. Studies involving these data are generally divided into two main categories:

• Forward Problem: predicting a spectrum based on molecular structure information. While spectroscopy devices can generate spectra from molecular samples, solving the Forward Problem (structure-to-spectrum problem) using AI models is highly valuable and offers several key advantages. First, it reduces the need for costly and time-consuming experimental measurements by enabling rapid spectral predictions. Second, it enhances the understanding of fundamental relationships between molecular structures and their spectral signatures. Such structure-to-spectrum correlation is crucial for scientists to know what molecule(s) are present for drug discovery, biomarker research, natural product synthesis, and other research areas [Mandelare et al., 2018]. Lastly, it expands applications beyond experimental limits. Some molecules are difficult to analyze using standard spectroscopy due to low concentrations, unstable intermediates, or extreme environmental conditions. AI solutions enables insights into such challenging cases where direct measurement is impractical.
• Inverse (Backward) Problem: deducing the molecular structure based on experimentally obtained spectra, also known as molecule elucidation, is a crucial task in life

sciences, chemical industries, and other fields [Sridharan et al., 2022a; Yao et al., 2023]. Resolving this problem enables researchers to identify unknown compounds, verify chemical compositions, and gain deeper insights into molecular behavior, ultimately advancing scientific discovery and industrial applications. However, molecular elucidation remains a time-consuming and complex process that heavily relies on human expertise. Identifying spectrum-to-structure correlation is particularly challenging, requiring analysts to distinguish real peaks and accurately deduce their chemical meaning. Manual interpretation is labor-intensive, has limited scalability, and is also prone to misinterpretation due to overlapping signals, sample impurities, and isomerization issues. This is where AI can play a transformative role, automating spectral interpretation, and accelerating the resolution of inverse problems.

Note that the above definition of the forward/inverse problem is in accordance with what is commonly referred to in the community [Lu et al., 2024]. However, the opposite definition exists in some contexts, e.g., in [Beck et al., 2024], where the inverse problem focuses on predicting spectra, while the forward problem refers to molecular deduction from given spectra. This difference in terminology highlights the slightly varying perspectives across disciplines and underscores the need for clear definitions when discussing these concepts in the context of spectroscopy and ML applications.

2.2 Roadmap of SpectraML

ML has revolutionized the way spectroscopic data is analyzed, offering new pathways to extract deeper insights, accelerate workflows, and uncover patterns beyond human capability. Historically, the use of computational techniques in spectroscopy was limited to basic pattern recognition and property prediction tasks [Elias et al., 2004]. This changed with the advent of deep learning and advanced ML frameworks that have enabled transformative capabilities across the entire spectrum analysis pipeline. For instance, CNNs excel in tasks such as peak detection [O'Shea, 2015] and deconvolution [Hu et al., 2024], akin to identifying features in an image, while RNNs and transformers [Schuster and Paliwal, 1997] handle sequential spectral data, similar to interpreting audio signals, making them suitable for reaction monitoring and dynamic studies. For example, CASCADE [Guan et al., 2021] accelerates the prediction of chemical shift in NMR spectra by 6000 times comparing to the fastest DFT method, enabling real-time NMR chemical shift predictions from simple molecular representations.

As spectroscopic datasets have grown in size and complexity, ML has demonstrated exceptional scalability and adaptability. The shift from early predictive models to modern generative and reasoning frameworks, such as attentionbased transformers and foundation models, has redefined the scope of spectral analysis. Generative models enable the simulation of spectra based on molecular structures [Goldman et al., 2023], addressing the forward problem, while reasoning-driven models tackle the inverse problem, predicting molecular structures with enhanced accuracy [Alberts et al., 2024a; Alberts et al., 2023]. More discussion regarding these two types of probelms is presented in next section. These developments have brought unprecedented precision and speed to applications ranging from molecular characterization to reaction pathway prediction. For example, IMPRESSION [Gerrard et al., 2020] predicts NMR parameters with near-quantum chemical accuracy while accelerating computational time from days to seconds.

3 SpectraMLMethodologies Summary

In this section, we present a detailed discussion of the machine learning methodologies that address the twin challenges: the forward (molecule-to-spectrum) and inverse (spectrum-to-molecule) problems. Our discussion is organized around four core components. We begin by examining the data representations and preprocessing strategies that serve as the foundation for effective spectral modeling. We then focus on the forward problem of predicting spectral signatures from molecular structures, followed by a discussion of the inverse problem of inferring molecular structures from spectral data. Finally, we describe emerging unified frameworks and cross-modal integration approaches that promise to address both challenges simultaneously. A summary of the discussed work is presented in Table 1.

3.1 Data Representations and Preprocessing

The quality of spectral analysis is fundamentally determined by how both molecular and spectral data are represented and preprocessed. In SpectraML, spectral data may be expressed as vectors, sequences, or images, while molecular structures are encoded using vector-based descriptors, simplified molecular input line entry system (SMILES) strings, 2D graphs, or 3D coordinates. Such diverse representations are essential for capturing the intricate details of molecular interactions. However, the high-dimensional and heterogeneous nature of spectral data, combined with challenges such as noise, baseline drift, and instrument variability, demands robust preprocessing pipelines. Early work demonstrated that conventional normalization and alignment techniques were insufficient for fully preserving the chemical information embedded in these datasets. Recent studies, including those by [Gastegger et al., 2017; Gerrard et al., 2020], have underscored the importance of integrating domain-specific knowledge, such as physics-informed normalization and tailored feature extraction into the preprocessing stage. More recent work, such as [Alberts et al., 2024a; Alberts et al., 2024b], is building a large-scale spectral dataset and harnessing the power of transformer-based models to map the latent representations of spectral data, thereby paving the way for robust and generalizable spectral analysis frameworks. These advances ensure that the learned representations are both resilient to experimental artifacts and chemically meaningful, thereby establishing a strong foundation for addressing both forward and inverse tasks.

3.2 Forward Problem: Molecule-to-Spectrum

The forward problem in SpectraML aims at predicting spectral information directly from known molecular structures, serving as an efficient alternative to computationally expensive quantum-chemical simulations and laborious experimental measurements. Forward-problem can also be extended to extract critical spectral features and related chemical properties. Therefore, the input of these ML-empowered solutions consists of molecules represented in different forms, such as SMILES strings, molecular graphs, or three-dimensional coordinates. The output can be either full spectra across different modalities (MS, NMR, IR, Raman, UV-Vis) or specific spectral features and chemical properties relevant to the target application. These ML approaches typically adopt an encoding-prediction framework, which predicts spectral features or related chemical properties in forms of regression or classification. In such architectures, the encoder transforms the molecular structure into a latent feature space that captures its essential chemical characteristics. The subsequent prediction stage then leverages this representation to predict partial spectra or specific spectral properties, depending on the target modality. While encoding and prediction are often implemented and trained end-to-end within a single model, without a strict separation (as demonstrated in example tasks D and E in Fig. 2), we structure the following discussion of related work based on the various forms of input and output involved in these problems, as they directly influence the selection and design of applicable machine learning models.

Input Encoding. As summarized in Table 1, the input to the forward problem is often in the form of vectorbased molecular features/descriptors, 2D, or 3D molecular graphs. This determines the choice of encoding, which is typically implemented as an MLP for vector-based molecular features/descriptors and a GNN for 2D and 3D molecular graphs. While vector representations are straightforward to handle, molecules represented as graphs require more sophisticated processing. Message-passing layers within GNNs effectively capture the structural and relational information between atoms and bonds. These graph-based encoders are typically paired with regression or classification modules to predict continuous properties, such as 1H and 13C chemical shifts [Guan et al., 2021; Kwon et al., 2020; Kang et al., 2020; Jonas and Kuhn, 2019], or to learn spectral features like excitation energies and spectral line shapes [McNaughton et al., 2023; Chen et al., 2022; Singh et al., 2022]. Alternatively, encoders may utilize direct coordinate-based features. For example, physics-informed neural networks extract vibrational properties directly from atomic coordinates-integrating experimental observations with quantum chemical insights-to predict key quantities such as dipole moment derivatives and polarizability tensors [Schütt et al., 2021; Gastegger et al., 2021; Chen et al., 2024; Sowa and Rossky, 2024].

Output Prediction. The "Task Type" and "Output" column in Table 1 for the forward problem indicate that the output prediction is mostly in the form of regression. For example in MS prediction, molecular substructures are mapped to fragment m/z values and intensities. In NMR spectroscopy, three-dimensional molecular graphs serve as inputs to predict continuous 1H and 13C chemical shifts, which in turn enables the accurate prediction of coupling constants and supports MAS-based spectral reconstructions [Cordova et al., 2023]. For IR, Raman, and UV spectroscopy, the prediction stage typically employs regression or classification layers to generate vibrational properties. In particular, key quantities-such as dipole moment derivatives and polarizability tensors-are predicted to capture the essential physical characteristics of the molecules [Schütt et al., 2021; Gastegger et al., 2021; Chen et al., 2024; Sowa and Rossky, 2024]. For ultraviolet and electronic spectra, excitation energies and spectral line shapes are predicted [McNaughton et al., 2023; Chen et al., 2022; Singh et al., 2022]. The prediction could also involve validating subformulas and predicting discrete spectral features [Goldman et al., 2023; Zhu et al., 2020; Young et al., 2024; Park et al., 2024; Zhu and Jonas, 2023; Murphy et al., 2023; Goldman et al., 2024], with some approaches further extending the framework to perform joint property prediction [Voronov et al., 2022]. The output can also take the form of a sequence of spectral tokens [Wei et al., 2019]. In this case, a generative model is employed to map a SMILES string to the full spectrum, enabling sequence-based spectral prediction.

Extension of Forward Problems. Beyond simulating spectral profiles, these forward-modeling approaches also support property-focused tasks: classification models can reveal reaction behaviors in MS-based reactivity analyses [Fine et al., 2020a], while hybrid ML-first-principles methods utilize IR data to infer adsorption energies and bond strengths [Du et al., 2023]. Additional efforts have extended these frameworks to predict other physicochemical parameters, such as logD values at various pH levels [Leniak et al., 2024], thereby supporting broader applications in drug discovery, catalyst design, and materials optimization.

3.3 Inverse Problem: Spectrum-to-Molecule

The inverse problem in spectral analysis aims at inferring a molecular structure directly from its measured spectrum, providing a complementary approach to traditional structure elucidation methods. In this task, the input consists of spectral measurements that can vary widelyfrom one-dimensional NMR signals and high-dimensional spectral vectors to imagelike two-dimensional matrices (e.g., from NMR or IR) and sequential data frommass spectrometry (MS). The output is the predicted molecular structure, commonly represented as a molecular graph or an SMILES string. ML approaches to the inverse problem typically adopt either an encoding-decoding scheme, where the spectral data is transformed into a latent representation and then decoded into a molecular structure (e.g., task B in Fig. 2), or an encoding-prediction framework, which directly predicts molecular substructures or functional groups from the spectral features (e.g., task C in Fig. 2). In such architectures, the encoder processes the input spectra to capture the critical information necessary for structure elucidation, and the subsequent decoder or classifier reconstructs the corresponding molecular representation.

Input Encoding. As presented in the bottom part of Table 1, the input to inverse modeling typically consists of one-dimensional 1H or 13C NMR spectra, which are signals along a single frequency axis, representing the chemical shift of the nuclei being observed, and often represented as high-dimensional vectors. For example, [Hu et al., 2024] employs a multitask, transformer-based model to encode 1D NMR signals into a latent space, facilitating the reconstruction of full molecular structures and substructure arrays. Similarly, [Huang et al., 2021] integrates convolutional neural networks with beam search to process spectral inputs, predicting substructure probabilities and iteratively assembling complete molecular graphs. In another approach, [Yao et al., 2023] leverages a bidirectional, auto-regressive transformer (BART) [Lewis, 2019] that is pre-trained on large-scale molecular data and fine-tuned with 13C NMR constraints. Additional methods, such as that of [Alberts et al., 2023], tokenize NMR spectral features into sequences for encoding, while [Sridharan et al., 2022b] combines Monte Carlo Tree Search with graph convolutional networks to iteratively build molecular graphs guided by spectral cues. These encoder designs are crucial for capturing both local and global spectral patterns that underpin accurate molecular reconstruction.

Output Decoding and Prediction. When the output of the inverse problem is a molecular structure, the task becomes a generative problem, where the decoder functions as a generator to reconstruct the molecular structure from spectral data, either as a sequence of tokens representing SMILES strings or by progressively constructing molecular graphs. For instance, [Alberts et al., 2023] utilizes a transformer decoder to convert tokenized NMR or IR spectra into SMILES strings, treating each spectral absorption value as a sequence element in a translation-like process. Similarly, [Jonas, 2019] frames the molecule reconstruction task as a Markov decision process (MDP) and incrementally reconstructs the molecule through a relation network. Moreover, additional constraints can be incorporated to refine the generative process; for example, [Sun et al., 2024] couples generative models with contrastive retrieval to enhance candidate matching accuracy, [Zheng et al., 2024] focuses on classifying seized substances from 1H and 13C NMR data, and [Tian et al., 2024] verifies proposed structures through joint analysis of image-like spectral data and graph-based molecular features.

Alternatively, an encoding-prediction framework maps the spectral representation to discrete structural elements, such as molecular substructures or functional groups, without generating an entire molecular structure. In this paradigm, the model deduces how atoms and functional groups are arranged to produce the observed spectral features, a capability that is particularly valuable for applications ranging from natural product identification to forensic analysis. For example, in MS context, MEDUSA [Boiko et al., 2022] and CANOPUS [Dührkop et al., 2020] incorporate classification and ranking layers to discriminate between candidate metabolites based on MS and MS/MS features, while CandyCrunch [Urban et al., 2024] predicts glycan topologies by analyzing tandemMS data. In IR spectroscopy, CANDIY-spectrum [Fine et al., 2020b] and CNN-based methods [Enders et al., 2021] focus on identifying diagnostic functional groups from characteristic absorption patterns. Similarly, transformer-based networks for NMR leverage 1H and 13C spectra to predict both key substructures and complete molecular formulas for robust classification and reconstruction [Huang et al., 2021; Hu et al., 2024; Alberts et al., 2023]. By directly extracting these critical features, the encoding-prediction paradigm offers an interpretable and efficient alternative to generative approaches for structure elucidation.

Extension of Inverse Problems. Beyond full structure elucidation, inverse SpectraML can be extended to recover detailed substructural information and functional group classifications that are essential for rapid compound identification and downstream analysis. For instance, sequence-tosequence models applied to MS/MS data-such as those demonstrated in Casanovo [Yilmaz et al., 2022]-and hybrid systems that combine substructure detection with full structure generation [Kim et al., 2023; Yao et al., 2023] further enhance compound identification when integrated with spectral databases, as seen in CFLS [Sun et al., 2024]. These extended approaches broaden the applicability of inverse SpectraML to diverse fields, from natural product discovery and metabolite screening to forensic investigations, thereby significantly reducing the reliance on time-intensive manual verification.

3.4 Unified Frameworks and Cross-Modal Integration

Recognizing that forward and inverse problems share common underlying chemical principles, recent research has begun to develop unified frameworks capable of addressing both tasks concurrently. Foundation models pre-trained on large, heterogeneous spectral datasets are at the forefront of this endeavor. These models leverage cross-domain learning to capture shared features across diverse modalities, such as IR, NMR, MS, and Raman-thereby enabling few-shot and zero-shot learning capabilities [Bommasani et al., 2021]. Concurrently, physics-informed generative models, including diffusion models and GAN-based super-resolution techniques, have been introduced to synthesize high-fidelity spectra while respecting known chemical constraints [Cordova et al., 2023]. Hybrid architectures that combine the relational modeling strength of GNNs with the sequence modeling capabilities of transformers offer a particularly promising route toward integrated spectrum analysis [Young et al., 2024]. Moreover, foundation models are steering the advancement of reasoning-driven spectrum analysis, particularly in complex inference tasks such as spectral deconvolution, peak assignment, and spectral consistency validation [Guo et al., 2024; Su et al., 2024; Alberts et al., 2024b]. These models can reason about ambiguous spectra by leveraging prior chemical knowledge to infer plausible molecular structures, resolve overlapping spectral features, and predict missing spectral regions. By unifying forward and inverse tasks within a single framework, these emerging approaches not only alleviate issues of data scarcity through synthetic data generation but also enhance model robustness and interpretability. This integrated perspective is poised to accelerate discovery in diverse domains such as drug development, materials science, and environmental monitoring.

4 Challenges and Opportunities

4.1 Data Quality, Scarcity, and Complexity

SpectraML faces several interrelated challenges arising from the inherent nature of experimental spectral data and the limitations of current ML approaches. First, the variability in data quality is a significant obstacle. Experimental spectra are often compromised by noise, baseline drifts, and instrument-toinstrument discrepancies, leading to inconsistencies in spectral resolution and intensity. Such variability complicates model training and can severely degrade the predictive performance of ML algorithms-especially when preprocessing pipelines are insufficiently robust. Moreover, the scarcity and imbalance of high-quality, annotated spectral datasets, particularly for rare or complex compounds, further exacerbate the issue. The limited availability of training data not only hinders the generalization of models across diverse chemical spaces but also increases the risk of overfitting, necessitating strategies such as data augmentation and transfer learning.

In addition to data quality challenges, the intrinsic complexity of spectral data presents a formidable hurdle. Spectral measurements typically exhibit high dimensionality and overlapping peaks, making feature extraction a nontrivial task. Current ML models often struggle to capture the nuanced, high-dimensional patterns inherent in such data, which leads to suboptimal performance in tasks like peak detection and feature discrimination. Furthermore, many existing architectures are not designed to fully leverage the domain knowledge embedded in spectroscopic data, thereby limiting their ability to exploit underlying chemical and physical principles.

A further challenge arises from the need to integrate data from multiple spectroscopic techniques (e.g., IR, MS, and NMR), each characterized by distinct scales, formats, and noise properties. Developing effective fusion strategies to reconcile these differences into a unified model is nontrivial. Most current ML architectures are optimized for singlemodality inputs and often fail to capture the critical crossdomain relationships needed for accurate spectral analysis. Additionally, achieving model interpretability-so as to provide meaningful insights into the underlying chemical phenomena-requires a careful balance between model complexity and transparency. Addressing these challenges is crucial for advancing the state of SpectraML and ensuring that ML-driven approaches can fully harness the rich information contained in spectral data.

4.2 Opportunities and Emerging Paradigms

Synthetic Data Generation and Physics-Informed Methods. To address challenges stemming from scarce and variable spectral data, AI-based generative models-such as large language models (LLMs) and diffusion models-have emerged as effective tools for advanced data augmentation [Luo et al., 2023; Bushuiev et al., 2024]. These models can learn complex, high-dimensional distributions from experimental spectra and subsequently generate synthetic data that mirrors key structural patterns and nonlinear dependencies [Alberts et al., 2024a; Alberts et al., 2024b]. Such synthetic spectra can supplement limited training sets, improving model robustness and generalization.

Incorporating physics-informed constraints-such as conservation laws, known intensity ratios, and chemical shift rules-into the generative process further enhances the realism and interpretability of the outputs. These priors guide the model toward producing chemically valid spectra, offering a compelling alternative to traditional simulation techniques like DFT or molecular dynamics. This hybrid strategy not only mitigates data scarcity but also reduces computational cost, facilitating faster iteration in applications ranging from materials science to pharmaceutical research.

Foundation Models: A New Paradigm for SpectraML. Foundation models [Bommasani et al., 2021; Moor et al., 2023], trained on large-scale, multimodal spectral datasets (e.g., IR, NMR, MS, Raman), offer a unified framework for spectral analysis. By leveraging cross-domain learning, they capture both global chemical patterns and local spectral signatures, supporting few-shot and zero-shot adaptation for diverse tasks such as peak prediction, spectral reconstruction, and structure inference.

These models integrate domain-specific priors directly into their architectures, enabling both forward and inverse reasoning. Their advanced inference capabilities allow them to resolve ambiguous or overlapping spectral signals through multi-step reasoning and modality fusion. Built-in mechanisms for uncertainty quantification and error detection enhance model reliability and interpretability. Overall, foundation models provide a flexible, data-efficient, and explainable approach to tackling both conventional and novel challenges in spectral learning.

5 Conclusion

The SpectraML establishes a crucial intersection between machine learning and spectroscopy. In this work, we provide a comprehensive overview of SpectraML and present a unified roadmap that traces methodologies across multiple spectroscopic techniques and categorizes key advancements in forward and inverse problems. To support future research, we highlight emerging trends such as generative modeling and foundation models and release an open-source repository. This survey serves as a valuable resource for researchers in both chemistry and AI, fostering interdisciplinary collaboration and driving innovation in spectral analysis.

Acknowledgments

This work was supported by the National Science Foundation under the NSF Center for Computer Assisted Synthesis (CCAS), grant number CHE-2202693.