---
title: 'LAPIS: Diverse Research Applications'
url: https://www.emergentmind.com/topics/lapis
type: topic
---

# LAPIS: Diverse Research Applications

LAPIS is not a single technical object but a recurrent name and acronym used across several distinct research domains. In arXiv literature it denotes, among other things, a pathogen-genomics web service, a semiconductor process and detector platform, an art-focused dataset for personalized image aesthetic assessment, language-model-centered reasoning and interface systems, a performance-portable compiler framework, and, in the capitalization variant *LapIS*, Laplace importance sampling. The term also appears in materials research through studies of lapis lazuli as a geological and cultural-heritage material. The usages are therefore polysemous rather than historically continuous.

## 1. Principal research meanings

Across the supplied literature, LAPIS functions as a domain-specific label whose expansion depends entirely on context.

| Expansion or referent | Domain | Representative paper |
|---|---|---|
| Lightweight API for Sequences | Genomic epidemiology, outbreak data services | [2206.01210] |
| LAPIS Semiconductor / LAPIS 0.2 µm FD-SOI process | SOI pixel detectors, X-ray and particle sensors | [1511.05224] |
| Leuven Art Personalized Image Set | Personalized image aesthetic assessment | [2504.07670] |
| Language Model-Augmented Police Investigation System | Legal reasoning for police investigations | [2407.20248] |
| Lightweight API Specification for Intelligent Systems | LLM-facing API description | [2602.18541] |
| MLIR-based compiler framework | Performance-portable scientific and ML compilation | [2509.25605] |
| LAPIS-SHRED | Sparse spatiotemporal reconstruction and forecasting | [2604.01216] |

A separate capitalization, **LapIS**, denotes **Laplace importance sampling** in work on randomized quasi-Monte Carlo [2309.11025]. Outside acronymic usage, **lapis** also appears in studies of **lapis lazuli**, where the term refers to the stone itself rather than to a computational framework [1608.07409].

## 2. Outbreak genomics and pathogen surveillance

In genomic epidemiology, LAPIS stands for **Lightweight API for Sequences**, described as a **fast REST API for very large open virus sequencing databases** [2206.01210]. It was developed in response to the data-access bottleneck exposed by the SARS-CoV-2 pandemic and the 2022 mpox outbreak, where laboratories generated millions of pathogen genomes but standard repositories and general-purpose systems were not optimized for recurrent surveillance questions such as filtering by date, country, lineage, amino-acid or nucleotide changes, stratifying counts by geography or time, and downloading cleaned aligned sequences.

The API distinguishes **aggregated endpoints** from **per-sample/data-download endpoints**. The principal aggregation endpoint, `aggregated`, returns either a total count or counts stratified by specified metadata fields; dedicated mutation-distribution endpoints include `nuc-mutations`, `aa-mutations`, `nuc-insertions`, and `aa-insertions`. Non-aggregated retrieval is provided by `details`, `nuc-sequence`, `nuc-sequence-aligned`, and `aa-sequence-aligned/{gene}`. LAPIS supports exact metadata filters, interval filters such as `dateFrom` and `dateTo`, lineage wildcards such as `XBB.1*`, insertion wildcards such as `ins:1000:AAT?`, advanced Boolean and threshold-style variant definitions such as `[2-of: 100A, 101T, 102G]`, and the special `maybe(...)` semantics for ambiguity-aware upper-bound queries.

A notable semantic detail concerns mutation proportions: ambiguous or unknown bases are excluded from the denominator. In the paper’s example, if 10 sequences contain 3 with `A5G`, 3 with reference `A`, and 4 with `N` at position 5, the reported proportion is
$$
\frac{3}{6}=0.5
$$
rather than $\frac{3}{10}=0.3$. This design reflects the system’s emphasis on outbreak-oriented interpretation rather than on generic database behavior.

Implementation is centered on a **new in-memory database engine**. The storage model is **column-wise**: for each aligned genomic or amino-acid position, LAPIS stores a string whose $i$-th character corresponds to sequence $i$. This allows positional mutation queries to inspect only relevant columns rather than full sequences. Insertions are handled separately through an **insertion store** that maps inserted sequence values at a reference position to the set of sequence IDs carrying them. Query-time evaluation occurs in memory; **PostgreSQL is used only as background storage**. Operational optimizations include **Zstd level 3** compression, **Redis** caching tied to data versions, and explicit version reporting through the `LAPIS-Data-Version` HTTP header and a `dataVersion` JSON field.

The production figures reported for the private SARS-CoV-2 deployment are unusually large for an interactive API. As of January 2023 the instance contained **more than 14.5 million SARS-CoV-2 sequences** and exceeded **400 GB**. Between **25 January and 4 February 2023** it processed **over 20 million requests** with **mean response time 411 ms**, **median response time 1 ms**, **72%** of responses served from cache, and **83%** of requests completed within **500 ms**. It is the backend for **genspectrum.org**, including **CoV-Spectrum** and the mpox dashboard, and the paper explicitly suggests that infrastructures such as **GenBank** could integrate LAPIS-like API functionality rather than forcing secondary systems to duplicate storage [2206.01210].

## 3. SOI process technology and monolithic pixel detectors

In detector electronics, LAPIS refers to **LAPIS Semiconductor Co., Ltd.** and related process technology for **0.2 µm fully-depleted SOI** monolithic pixel detectors [1511.05224]. The core device concept is a monolithic SOI detector in which the **readout/electrical circuitry** is fabricated in the **top silicon layer**, while the **X-ray sensing p–n junction diode** is formed in the **high-resistivity handle wafer**. This supports a single-substrate architecture with a depleted sensing volume beneath integrated CMOS.

The process literature emphasizes several detector-specific innovations. A **buried P-well (BPW)** biased with **GND** suppresses back-gate effects; the paper reports that with BPW there is **“There is no Id-Vg change even with Vback = 100 V”** [1511.05224]. A **double-SOI wafer** introduces a **middle-SOI layer** beneath the transistors to compensate BOX charging effects and reduce crosstalk; with negative middle-SOI bias, IV characteristics become nearly pre-irradiation even after exposure up to **10 Mrad**. For **high-resistivity FZ-SOI**, slow ramp-up and ramp-down thermal recipes reduce slip generation to an **“acceptance level for pixel sensor fabrication”**. Backside processing with **wet etching after back grinding** and **laser anneal** removes a **0.3 µm-thick damaged layer** and suppresses abnormal leakage current under full depletion. For large-area imagers, **stitching exposure** yields chips of about **66 mm width** and **30 mm height**, with **14 µm** boundary area, pattern thinning **less than 0.1 µm**, and about **0.015 µm** Y-direction alignment.

A related study analyzes shielding mechanisms for **LAPIS-fabricated SOI pixel detectors** in counting-type pixels [1507.05394]. The problem is circuit-to-sensor **charge injection** and crosstalk. With a coupling capacitance of
$$
0.2~\mathrm{fF}/\mu\mathrm{m}^2
$$
a \(1~\mathrm{V}\) step on a \(1~\mu\mathrm{m}^2\) area induces about \(0.2~\mathrm{fC}\), or roughly **1200 \(e^-\)**, equivalent to about a **4.3 keV X-ray quantum**. Prototype chips **CPIXTEG3** and **CPIXTEG3b** compared **Nested-well** and **Double-SOI** shielding. The paper concludes that a shielding layer is indispensable for counting-type pixels and that **Double-SOI is superior to Nested-well in terms of shielding effectiveness and design flexibility**; in the Nested-well implementation the counter produced spurious signals corresponding to about **20 ke\(^{-}\)**, whereas in the Double-SOI version the authors report **no discernable crosstalk correlated to any clock edge**.

The same foundry platform underlies integration-type sensors such as **INTPIX4** [1804.03338]. That detector uses the **0.2 um low-leakage fully depleted (FD)-SOI CMOS process of Lapis Semiconductor Co., Ltd.**, with **1 poly, 5 metal layers**, **MIM capacitor and DMOS options**, **Core / I/O voltages of 1.8 V / 3.3 V**, and **8-inch SOI wafers**. INTPIX4 has **512 × 832** pixels, **17 × 17 µm\(^2\)** pitch, **13 parallel output lines**, and per-pixel **correlated double sampling (CDS)**. The transition from Cz to FZ substrates after 2011 reduced leakage current and enabled a thicker sensing depth at moderate bias; reported FZ performance includes detector gain **13.1 µV/e**, **FWHM = 624 electrons**, and **12.8% at 17.48 keV**.

A later LAPIS-based prototype for collider tracking is **CPV3**, fabricated in the **LAPIS 200 nm SOI process** and focused on a **pinned depleted diode (PDD)** with AC-coupling on the diode [2104.04936]. The design targets very small pixels, **16 × 20 µm\(^2\)**, with about **\(-60~\mathrm{V}\)** backside bias on a **310 µm** sensor. The study reports **very low leakage current**, **\(13~\mu\mathrm{V}/e^-\)** charge-voltage factor, raw equivalent input capacitance around **12 fF** for one structure, and about **3.5 fF of equivalent input capacitance** for the optimized in-pixel PDD capacitance. Among the three PDD variants, **PDD3** showed the lowest equivalent diode capacitance and better charge sharing. Taken together, these papers situate LAPIS as a detector-oriented SOI process line spanning X-ray imaging, counting pixels, and collider-grade monolithic sensor R&D [1511.05224].

## 4. Machine-learning datasets and latent spatiotemporal inference

In computational aesthetics, LAPIS denotes the **Leuven Art Personalized Image Set**, introduced as a dataset for **personalized image aesthetic assessment (PIAA)** and described as the **first dataset with images of artworks that is suitable for PIAA** [2504.07670]. It contains **11,723 images** sourced from **WikiArt**, mostly paintings with some sketches, spanning **26 styles** and **7 genres**. The dataset includes **7,976 figurative works** and **3,747 abstract works**. Annotation was conducted in an online study with **552 participants** recruited through **Prolific**; after quality control, each image had ratings from **an average of 24 annotators**. LAPIS provides both per-image aesthetic ratings and rich side information: personal attributes such as age, nationality, gender, education level, colorblindness, art interest, and art familiarity, plus image attributes including style, genre, and **31 image attributes** expanding to **47 image features**.

The benchmarking results are notable because they expose a large protocol gap. Under a traditional train/test split, **PIAA-MIR** and **PIAA-ICI** achieve **SROCC 0.6958** and **0.6941** on LAPIS, but under the stricter **4-fold cross-validation with disjoint train/test users** they drop to **\(0.2793 \pm 0.0215\)** and **\(0.2773 \pm 0.0235\)**. The ablation study shows strong sensitivity to **Art interest** and to **Style and genre**, while removing **Objective image attributes** slightly improves the baseline SROCC from **0.69583** to **0.70118**. The paper interprets this as evidence that current PIAA models do not generalize well to unseen users and that artistic image aesthetics remain difficult, especially for abstract or less-preferred genres.

In a different machine-learning setting, **LAPIS-SHRED** stands for **LAtent Phase Inference from Short time sequences using SHallow REcurrent Decoders** [2604.01216]. It addresses reconstruction and forecasting of complete spatiotemporal dynamics from **hyper-sparse sensor observations** confined to short temporal windows. The framework uses a **three-stage pipeline**: a SHRED model is pre-trained on simulation data to map sensor time-histories into a latent space; a temporal sequence model is trained on simulation-derived latent trajectories to propagate latent states forward or backward; and, at deployment, a short real observation window is encoded and expanded into a full latent trajectory before decoding back to the physical field. The architecture supports **bidirectional inference**, multiscale backbones, and even **single-frame terminal inputs** through static padding.

The empirical scope is broad: the paper reports **six experiments** spanning **turbulent flows, multiscale propulsion physics, volatile combustion transients, and satellite-derived environmental fields**. The summary table gives **NRMSE 0.046** for **2D KS**, **0.044** on velocity and **0.036** on vorticity for **2D KF**, **0.037** forward and **0.033** backward for **2D KVS**, **0.114** for **HF-RDE**, **0.025** on pressure and **0.032** on temperature for **1D RDE ignition**, and **0.167** forward and **0.130** backward for **NDSI**. The paper states that full trajectories can be reconstructed from windows spanning only **\(7\%-20\%\)** of the temporal domain using as few as **3 sensors**, often with NRMSE below **5%**. This suggests a conceptually distinct but similarly named LAPIS: a latent-trajectory inference layer built on top of SHRED rather than a dataset [2604.01216].

## 5. Language-model and API-oriented systems

In legal NLP, LAPIS expands to **Language Model-Augmented Police Investigation System**, an AI decision-support system for police officers conducting criminal investigations in Korean [2407.20248]. The system formalizes a crime-investigation legal reasoning task in which, given a hypothesis \(h\) and investigative context \(\mathcal{C}\), an evaluator model returns both a binary assessment and a rationale. Retrieval is integrated through a **Crime Investigation KnowledgeBase (CIKB)**, and deployment follows the pattern
$$
\mathbf{y} = f(h, \mathcal{C}, \mathbb{P}) \text{ where } \mathbb{P} = g(h, \mathcal{C}).
$$
The dataset, **Crime Investigation Legal Reasoning (CIRL)**, was built from the Korean National Police Agency’s annual **Criminal Law Proficiency Exam** from **2013 to 2023**. The authors collected **1,760** original exam questions and converted them into **4,895** hypothesis instances, then expanded the rationale-bearing training set to **14,105** instances with GPT-4 generation and domain-expert curation by **20 law enforcement officials affiliated with the Korean National Police Agency**.

CIKB contains **19,388 dense embeddings**, derived from **687 paragraphs** of criminal investigation textbooks, **3,304 paragraphs** of criminal law articles, and **15,397 paragraphs** of criminal court rulings. At inference time the system retrieves the **top 5 relevant premises**. The best evaluator is **EEVE-Korean-Instruct-10.8B-v1.0**, fine-tuned for **5 epochs** with **Paged AdamW**, **learning rate 0.0001**, and **QLoRA** on **two NVIDIA A100 80GB Tensor Core GPUs** for **approximately 30 hours**. On the held-out test set, the reported total performance is **ACC 0.74, F1 0.79**, slightly above GPT-4’s **0.73 / 0.79** overall, with especially strong gains on the **Crime Investigation** category. The paper’s qualitative analysis emphasizes that LAPIS grounds rationales in premises and may avoid failures such as GPT-4 citing a **non-existent precedent**.

A different LLM-centered use of the name is **LAPIS: Lightweight API Specification for Intelligent Systems** [2602.18541]. Here LAPIS is a domain-specific specification format designed for **LLM consumption**, motivated by the claim that **OpenAPI is the wrong optimization target for LLMs**. The document structure has up to seven fixed-order sections: `[meta]`, `[types]`, `[ops]`, `[webhooks]`, `[errors]`, `[limits]`, and `[flows]`, with `[meta]` and `[ops]` required. The design centralizes repeated information, represents operations as signature-like entries, gives first-class structure to webhook triggers and rate limits, and allows operation-flow declarations that OpenAPI either duplicates redundantly or cannot represent structurally.

The token-efficiency claims are substantial. Against five production API specifications—**GitHub (1,080 endpoints)**, **DigitalOcean (545 endpoints)**, **Twilio (197 endpoints)**, **HTTPBin**, and **Petstore**—the paper reports an average reduction of **85.5%** versus OpenAPI YAML and **88.6%** versus OpenAPI JSON using the `cl100k_base` tokenizer. For GitHub, the count falls from **1,811,843** YAML tokens to **313,101** LAPIS tokens; for DigitalOcean, from **586,731** to **54,201**; and for Twilio, from **306,453** to **24,197**. The paper also highlights structural waste in OpenAPI, noting that GitHub’s spec contains **1,594** error definitions across **1,080** operations but only **14 unique error codes**, with **404** repeated **531** times. In this meaning, LAPIS is a reasoning-optimized projection of OpenAPI rather than an operational API service [2602.18541].

## 6. Compiler infrastructure and other scientific usages

In programming languages and systems, LAPIS is also the name of an **MLIR-based compiler framework** for scientific computing and machine learning [2509.25605]. The stated goal is simultaneous **performance**, **portability**, and **productivity**, particularly where computational science and machine learning workflows converge. The compiler lowers **linalg-on-tensors** MLIR through a custom **Kokkos dialect** and emits standalone **Kokkos C++**. It introduces lowering passes such as `linalg-to-kokkoskernels`, `dense-linalg-to-parallel-loops`, `kokkos-loop-mapping`, and `kokkos-dualview-management`, and supports both dense and sparse linear algebra as well as interoperability between **PyTorch** and **Kokkos**.

The dense GEMM benchmark is numerically explicit. For **4096 × 4096 FP32** SGEMM, LAPIS and KokkosKernels are effectively identical: **1.83 ms** versus **1.86 ms** on **Granite Rapids**, **0.359 ms** versus **0.358 ms** on **H100**, and **0.264 ms** versus **0.263 ms** on **MI300A**. For SpMV, the paper reports performance similar to handwritten **Kokkos Kernels**, somewhat slower than **cuSPARSE**, similar to **MKL** on Granite Rapids, and similar to **rocSPARSE** on MI300A. It also demonstrates model integration for **ResNet18** and a SciML surrogate in **MALA**, with MALA inference for **8748 atoms** taking about **0.5 ms**. In this usage, LAPIS is a compiler and runtime-emission framework rather than an API or dataset.

A mathematically separate capitalization, **LapIS**, refers to **Laplace importance sampling** in work on importance sampling combined with randomized quasi-Monte Carlo [2309.11025]. The proposal is Gaussian with mean at the mode of the log-target and covariance given by the inverse negative Hessian:
$$
\bm \Sigma_{\star}=(-\nabla^2 H(\bm z_\star))^{-1}.
$$
Under a randomly shifted rank-1 lattice rule and suitable weighted-RKHS conditions, the paper proves the RMSE bound
$$
\sqrt{\mathbb{E}|(\hat{I}_N(G_{IS})-C|^2}= O(N^{-1+1/\alpha^2+\delta}), \quad \delta>0.
$$
For **LapIS**, sufficient conditions include both the growth condition
$$
g(\bm z)=O(\|\bm z\|^\beta),\qquad \beta<2,
$$
and the additional eigenvalue condition
$$
-\min_{i \in 1:d} h_i \max_{j \in 1:d} \lambda_j< \frac{1}{\alpha}.
$$
The paper explicitly remarks that the convergence conditions for LapIS are more strict than for **ODIS**, so the term here denotes a specific importance density rather than a standalone system.

Outside acronymic usage, **lapis** remains a materials-science and heritage-science term through **lapis lazuli**. A provenance study based on **CL**, **SEM-EDS**, **micro-Raman**, **IL**, **PIXE**, and **PIGE** proposes source criteria for material from **Afghanistan**, **Pamir Mountains in Tajikistan**, **Chile**, and **Siberia** [1608.07409]. Chilean samples are associated with **wollastonite** and strong bands at **~560 nm** and **~620 nm**; Pamir material with a **~690 nm** diopside band and cancrinite-group UV luminescence; Siberian material with **Ba > 1%** and **Sr ~6000 ppm**. A later colorimetric microscopy study treats lapis lazuli as a heterogeneous optical material composed primarily of **blue lazurite, yellow pyrite, and white calcite** [2304.07293]. For the polished lapis sample, the reported mean dominant wavelength is
$$
\bar{\lambda}_d = 448 \text{ nm}
$$
with standard deviation **12 nm**, and the mean excitation purity is
$$
\bar{p} = 0.527
$$
with standard deviation **0.069**. These non-acronymic usages underscore that LAPIS can signify either an institutional or algorithmic acronym or the mineral term itself, depending on disciplinary context.

Source: https://www.emergentmind.com/topics/lapis