Material quality control in industry and laboratories is gaining a new ally: generative artificial intelligence that acts as a "virtual spectrometer." This tool is SpectroGen, a new approach that, based on a spectrum measured with one technique (e.g., infrared), realistically and with very high accuracy generates what the spectra would look like in other techniques (e.g., X-ray diffraction or Raman spectroscopy). This dramatically shortens and cheapens a series of steps that until yesterday required multiple expensive and bulky instruments, specialized teams, and hours to days of measurement.

Why the "virtual spectrometer" is a game-changer

In the development of new batteries, faster electronics, or more effective drugs, two phases are critical: finding promising materials and verifying their structure and properties. While machine learning algorithms have proven excellent in searching material databases in recent years, the validation phase – the actual measurements using various spectroscopic and diffraction methods – has remained a bottleneck. Each technique reveals a different "dimension" of the material: infrared (IR) spectroscopy highlights functional groups, Raman vibrational states, and X-ray diffraction (XRD) the crystal lattice. In practice, this means multiple separate instruments, separate sample preparations, and significant time consumption.

SpectroGen introduces a different workflow: once a sample is captured with a cheaper and more compact sensor (e.g., an IR camera on a production line), the generative model creates "equivalent" spectral signatures in other modalities. Operators then have enough information to make quality control decisions, without mandatory physical diffraction or secondary measurement on another instrument. In many cases, this will mean faster shipment of a product batch, less downtime, and over time – lower capital equipment costs.

How SpectroGen derives other spectra from just one measurement point

The classic approach to modeling spectra is "bottom-up": from atoms, bonds, and crystallography to the expected signals in a particular spectroscopic method. This path is computationally demanding and often inapplicable even for a single complex material. The team behind SpectroGen chose the opposite direction. Instead of the algorithm looking at chemical bonds, it observes the mathematical curves that make up the spectrum – distributions like Lorentzian and Gaussian, peak ranges, widths, and positions. Since different techniques have characteristic signal shapes (e.g., Raman often shows more "Gaussian" peaks, while IR prefers "Lorentzian" ones), the model learns the rules for translating one "alphabet" of curves into another.

This physics-guided, yet generative logic makes the tool robust beyond a narrow set of samples. In practice, the model first "adopts" the features of the spectral curves of the observed modality and then, using the learned correspondences, proposes how this same information is "mapped" into the target technique. The key difference is that we are not talking about a naive mapping or simple statistical interpolation, but about learning physically meaningful transformations between spectral spaces.

Speed and accuracy: from hours and days to a minute

In typical material testing workflows, preparation, measurement, and analysis in XRD or Raman can take hours, and in more complex validations, even longer. SpectroGen shortens this step to less than a minute per sample – fast enough to keep pace with modern production lines. What makes it useful in industry is not just its speed, but also the precision of its prediction: the high correlation between the generated and the actually measured spectrum in the target technique shows that the tool does not invent artifacts, but transfers information that is relevant for quality decisions.

This opens up practical scenarios like "one sensor – multiple insights." A factory can continuously use an infrared scanner on incoming raw material or a finished electrode, and then in the software get a reliable "proxy" XRD diffractogram to check the phase composition or a Raman spectrum to look for impurities and by-products. It is not realistic to expect that every case will completely eliminate physical measurements, but it is likely that their frequency and the need for expensive equipment will significantly decrease.

What this means for key industries: batteries, semiconductors, and pharmaceuticals

Battery manufacturing. Controlling the phase purity of cathode and anode materials (e.g., NMC, LFP, graphite, silicon) and monitoring degradation products are crucial for performance and safety. IR and Raman are already standard tools for detecting functional groups, organic binders, and the SEI layer, while XRD provides insight into the crystal structure and any secondary phases. The combination of "IR measurement + SpectroGen" can allow manufacturers to perform quick "XRD-like" checks on every batch, reserving physical diffraction for occasional calibration or disputed cases.

Semiconductors and electronics. In epitaxy, thin-film deposition, and lithography, small variations in crystallinity and stress mean the difference between a working and a faulty chip. Generating diffractograms or Raman profiles from in-line optical measurements could detect deviations earlier and prevent waste – which is crucial in facilities with high OEE requirements.

Pharmaceutical industry. The identification of polymorphic forms, hydration states, and impurities often relies on a combination of Raman/IR/XRD. If routine work can rely on cheaper optical probes and digitally "conjure up" supplementary signals, batches can be qualified faster, and lot release procedures can be freed from expensive instruments and expert teams.

"Physics-in-the-loop": why it matters

Many generative models suffer from the problem of "hallucinations" – plausible but incorrect results. SpectroGen bypasses this trap by explicitly incorporating physical intuition about the shape of spectral lines and noise, as well as the limitations that apply in real instruments. Such "physics-in-the-loop" means that the model will not, for example, generate impossibly sharp peaks where broader signals are expected due to the instrument function or the nature of the transition. This is also essential for auditability: experts can understand why the model created a certain pattern and where its limits are, which helps in regulated industries.

From research breakthrough to production line

The first major test for SpectroGen was conducted on rich sets of spectral data for minerals, covering X-ray diffraction, Raman, and IR. Such repositories, built over years, are ideal for learning the relationships between different modalities because they link chemical formulas, crystallographic parameters, and measurement signatures. After training on a subset of samples, the model was tested on "new" samples and showed a very high agreement between the generated and actual spectra in the target techniques. This result is an important indicator of generalization, as the mineral database covers a wide spectrum of structures and compositions.

The transfer to industry will require two additional steps. The first is process calibration: the model is fine-tuned to your equipment, conditions, and goals (e.g., detecting impurities in parts per thousand or detecting a specific polymorph). The second is validation in a real-world environment – establishing decision thresholds, defining a protocol for when a generated spectrum and rapid analysis detect anomalies, and setting a rhythm for occasional physical measurements for verification and maintaining traceability.

How to integrate it into an existing laboratory or factory

• "One modality as the source of truth": Choose the most accessible and robust sensor (in practice, this is often FTIR or NIR). This instrument becomes the continuous source of data that SpectroGen then translates into other "virtual" modalities.


• Gradual implementation: Initially, confirm every critical batch with physical measurements in another technique. As confidence in the model increases, the frequency of these confirmations can be reduced with clearly defined control points.


• Digital traceability: Store each generated spectrum with its corresponding "source" spectrum and sample metadata. This creates an archive useful for audits, internal scientific control, and future re-training.


• Limits of applicability: If a material's signal characteristics deviate from what the model was trained on, the training set needs to be expanded. Otherwise, the model may produce consistent but biased projections.

Use case examples with specific metrics

Line clearance and recipe change: When switching suppliers of an active substance or moving to a new batch of cathode powder, introducing SpectroGen allows for the comparison of "virtual XRD" and "virtual Raman" signatures with a reference database. If the deviations are below a set threshold, the line continues without interruption; otherwise, the sample is sent for physical diffraction. This reduces the "waiting time for the instrument" from hours to minutes.

Monitoring aging and degradation: Generated spectra can highlight the appearance of secondary phases or changes in vibrational modes typical of degradation. In batteries, this is seen through peak shifts; in pharmaceuticals, through the formation of an undesirable polymorph. In both cases, early detection saves resources and prevents waste.

Energy and maintenance management: If some physical measurements are replaced by digital ones, key instruments consume fewer resources, are stopped less frequently, and remain in calibration longer. This reduces the cost per sample and frees up expert staff.

Technical insight: what's "under the hood"

The model for transferring between spectral modalities combines several ideas. First, spectra are not treated as simple arrays of values, but as superpositions of parametric functions (e.g., a set of Gaussian and Lorentzian peaks with limited widths and noise), which requires the generated output to respect physical constraints. Second, training is conducted on large datasets that include pairs (or triplets) of recordings of the same sample using different techniques, thereby learning robust mappings. Third, so-called physical priors can be incorporated into the model – e.g., the expected signal-to-noise ratio, the maximum number of significant peaks, and typical instrument functions – which reduce the risk of "realistic but inaccurate" synthesis.

It is important to note that the tool is not limited to minerals. Spectroscopy is universal: wherever light (or radiation) interacts with matter and leaves a measurable trace, there is potential for transfer between modalities. This is precisely why development is moving towards adaptation for biomedical diagnostics, food and water safety monitoring, and environmental and agricultural surveillance.

Comparison with traditional approaches and other AI tools

Unlike models that "guess" the structure from one method and then "predict" what another would look like, SpectroGen avoids direct chemical or crystallographic backtracking. The advantage lies in its robustness and speed: it skips the interpretation layer, which is often sensitive to noise and uncertainty. Compared to pure deep networks without physical constraints, the approach with embedded priors shows better generalization and less susceptibility to artifacts.

Where digital twins of equipment and process monitoring software are already in use, SpectroGen can be added as a module for spectral "cross-modality" transfer. This progressively creates a closed loop: inexpensive optical measurements provide rich digital signals, a generative model translates them into the measurement "languages" that engineers need, and decisions are fed back into the process.

What is needed to get started: data, people, and standards

Data: The greatest benefit is achieved if an organization already has an archive of spectral pairs (e.g., IR and XRD of the same samples). If the archive is heterogeneous, it needs to be cleaned and the metadata (samples, lots, instrument, conditions) consistently labeled. For new environments, it is recommended to purposefully collect several hundred well-documented pairs to ensure a solid starting point.

People: Although the user interface is simple, the greatest value comes when process engineers, analysts, and IT/OT teams work together. Spectroscopists help set quality rules and thresholds; data scientists take care of training and validation; the engineering team is responsible for integration with the Manufacturing Execution System (MES) and Laboratory Information Management Systems (LIMS).

Standards and compliance: In pharmaceuticals and related regulated areas, it is necessary to define protocols for model verification and re-validation (e.g., a periodic "challenge set" with physical measurements). Transparency about the model's limitations, versioning, and input/output traceability are crucial for inspections.

From the lab to the field: diagnostics and agriculture

Development doesn't stop at materials for batteries and semiconductors. The same "one measurement – multiple insights" concept holds promise in rapid medical tests, where the goal is to obtain a reliable spectroscopic signature from a drop of blood or tissue. In agriculture, optical probes on drones or tractors collect large amounts of NIR/hyperspectral data; converting these signals into "virtual" Raman or other modalities allows for more detailed control of crop health, stress, and the presence of pathogens in the field.

Practical questions and common observations from pilot projects

• "What if my sample is outside of anything the model has seen?" If it's a completely new type of material or extreme conditions, the training set needs to be specifically supplemented. In the meantime, the model can still serve as a quick filter, and "critical" samples are sent for physical measurements.


• "How often should physical confirmations be done?" In the initial phase, it is recommended to confirm every key decision with a physical XRD/Raman measurement. After a few weeks or months, when statistics show stability, the frequency of confirmations can be reduced to periodic samples.


• "Can the model replace all instruments?" No – its role is to reduce the frequency and shorten the time to a decision. Physical instruments remain the reference for calibration and disputed cases.


• "How does the tool handle noise and bad samples?" Since instrument functions and typical noise levels are built into the model, the generated spectrum will not "polish" the data beyond recognition. If the input spectrum is poor (e.g., saturated or with too many artifacts), the system can flag suspicion and request a re-acquisition.

The bigger picture: AI that understands science, not just data

The significance of SpectroGen goes beyond a single tool – it shows how generative AI can be applied responsibly and usefully in materials science when "enriched" with physical priors. Instead of generating attractive but random results, the starting point here is the structure of the signal and the limits set by the instrumentation. Such an approach is increasingly attractive in disciplines where large measurement archives are available, but the time and cost of new acquisitions are the biggest obstacle.

Resources, communities, and next steps for teams who want to try it

Companies and laboratories considering a pilot implementation should inventory their existing spectroscopic archives, select critical material lines, and form a minimum validation set. Then, set internal success metrics: what is the acceptable difference between the generated and the real spectrum, what are the KPIs (e.g., reduction in batch qualification time, number of avoided re-measurements, savings in capacity on expensive instruments). In rare cases where a wrong decision is very costly, a more conservative threshold and a slower reduction of physical confirmations are recommended.

Implications for jobs and competencies

The automation of analysis does not mean fewer roles for people, but a different focus. Spectroscopy experts will be more involved in setting rules, designing experiments, and controlling model validity, while routine comparisons are taken over by digital tools. For data scientists, this opens up space for developing and maintaining specific models per product portfolio. Effective collaboration between these profiles is key to extracting real value from generative models.

A look towards 2026: personalized models by industry

As databases of spectral pairs expand and become standardized, the emergence of "vertical" models based on the SpectroGen principle is expected – especially for battery materials, pharmaceutical substances, and high-tech alloys. Such models will be able to offer ready-made, industry-specific settings, decision thresholds, and interfaces integrated with existing quality control systems. In addition, advances in compact optoelectronics will further lower the barrier to in-line implementation.

When "good enough" is truly good enough

In industrial practice, perfection is never sought, but rather a sufficiently good signal fidelity for an accurate and quick decision. In this sense, the virtual spectrometer has a clear place: on 9 out of 10 samples, it allows for qualification in a minute and frees up expensive resources; on the remaining sample that "whistles" – a message is sent to the laboratory, a physical measurement is made, and the model is updated if necessary. This rhythm is what many factories are missing today: predictable, scalable, and financially sustainable.

What procurement teams and management need to know

The economics of introducing a virtual spectrometer are not just about the cost of a software license. One must also consider the reduced load on existing instruments, the postponement or reduction of investment in new equipment, faster batch releases, lower operational risk of downtime, and a reduced number of man-hours on repetitive analyses. Even in a conservative scenario, the ROI is achieved through a combination of faster cycles and lower capital expenditures. In more advanced scenarios, when virtual modalities start being used for predictive process maintenance, the benefits grow exponentially.

Frequently asked questions in audits and regulatory reviews

• Model change documentation: Each new version of the model must have a list of changes, a regression test set, and a record of its impact on quality KPIs.


• Decision traceability: For each lot, record the input spectrum, the generated virtual spectra, the automated assessment, and the human decision. This allows for the reconstruction of the entire sequence of events.


• Bias detection plan: Periodically test the model on samples outside the usual range to detect systematic deviations.

Note on dates and technology development up to October 15, 2025

In the current timeframe up to October 15, 2025, SpectroGen has undergone a series of public presentations and technical reports and is being actively adapted for various industrial domains, including materials monitoring, biomedical diagnostics, and agriculture. At the same time, the communities around Raman, IR, and XRD reference databases continue to update their collections, which further improves the training and validation of the model in various applications.

Who will find this text particularly useful

Quality managers in battery and pharmaceutical manufacturing, laboratory managers, process engineers in advanced manufacturing, R&D department heads, and data scientists who want to reduce the number of expensive physical measurements while increasing the frequency and scope of checks. If you have existing pairs of spectra or the ability to collect a targeted validation series in a short time, the transition to "virtual modalities" can begin in the first phase without major organizational changes.

For those who want to delve deeper into the topic, a review of modern practices in generative artificial intelligence in materials science is useful, as are the current project pages and laboratories dedicated to spectroscopy and reference databases. Familiarize yourself with the basics of Raman spectroscopy, X-ray diffraction, and infrared spectroscopy, and with the possibilities of mapping them into digital twins of processes. Keep in mind that the key to success lies in careful validation on your own samples and setting clear decision thresholds – those that best align with the goals of your production line or laboratory.

Note on language and spelling: All terms and instrument names in the text are standardized in the Croatian professional language (Ramanova spektroskopija, rendgenska difrakcija/XRD, infracrvena/IR), while technology names are transcribed in their original form where common (SpectroGen, AI, generativni model).