---
title: 'Surya: Multi-Domain Disambiguation & Applications'
url: https://www.emergentmind.com/topics/surya
type: topic
---

# Surya: Multi-Domain Disambiguation & Applications

In arXiv usage, **Surya** is not a single scientific object but a term with several distinct referents. It denotes an **open-source, multilingual OCR toolkit** for printed text, a **366M parameter foundation model for heliophysics**, the **surname of multiple authors** in deep learning, causal set theory, stochastic processes, combinatorics, mathematical finance, and mathematical relativity, and, in a separate computer-vision context, the Hindu Sun God referenced through **Surya Namaskar**. The term therefore requires explicit disambiguation at the level of domain, paper, and technical context [2507.18264] [2508.14112] [1802.09979] [1311.2148] [2209.02492].

## 1. Referential scope in the literature

A recurring source of ambiguity is that identical surface forms correspond to fundamentally different entities.

| Referent of “Surya” | Domain | Defining description |
|---|---|---|
| Surya OCR | Document AI / OCR | Open-source, multilingual OCR toolkit for printed text |
| Surya foundation model | Heliophysics / foundation models | 366M parameter model trained on SDO observations |
| Surya Ganguli / Sumati Surya / Budhi Surya | Author surname | Appears in theory papers across multiple fields |
| Surya in Surya Namaskar | Yoga / computer vision | Religious-cultural referent inside pose-recognition work |

This disambiguation matters methodologically. In some papers, **Surya** names a concrete model family with architecture, training protocol, and benchmark results. In others, it is only an author surname and does **not** refer to an algorithm, framework, or module. The deep-learning paper “The Emergence of Spectral Universality in Deep Networks” states this explicitly: “Surya” there is **Surya Ganguli**, not a model named Surya [1802.09979].

## 2. Surya as an OCR and document-processing system

In “Zero-shot OCR Accuracy of Low-Resourced Languages: A Comparative Analysis on Sinhala and Tamil,” Surya is described as an **open-source, multilingual OCR toolkit** designed to support **90+ languages** and to benchmark well against cloud OCR services. Its printed-text OCR stack includes **line-level text detection**, **layout analysis**, **reading-order detection**, **table recognition**, and **LaTeX OCR**. Technically, its text detection was trained on **four A6000 GPUs for three days** using a diverse image set and is built on a **modified EfficientViT** architecture for segmentation, while its recognition model was trained on the same hardware for **two weeks** using a **modified Donut** model with **Grouped Query Attention (GQA)**, a **Mixture of Experts (MoE)** layer, UTF-16 decoding, and layer-configuration changes. The system is explicitly intended for **printed text**, not handwriting [2507.18264].

The same paper evaluates Surya in a **zero-shot OCR setting** on **Sinhala and Tamil**, where “zero-shot” means that the system was used as-is and was not fine-tuned on those languages for the study. The evaluation uses synthetic datasets with clean **black text on white backgrounds**: **6,969 image-text pairs** for Sinhala from `sinhala_synthetic_ocr-large`, and **7,000 samples** for a new synthetic Tamil dataset rendered using **six Tamil fonts** after filtering OPUS/OpenSubtitles Tamil lines longer than 40 characters. Accuracy is measured with **CER**, **WER**, **BLEU**, **ANLS**, and **METEOR**, with CER and WER defined as
$$
\text{CER} = \frac{S+I+D}{N}, \qquad \text{WER} = \frac{S+I+D}{N},
$$
and ANLS given through thresholded normalized Levenshtein similarity with threshold \(T = 0.5\) [2507.18264].

On **Sinhala**, Surya is the best system in the study across all metrics: **CER 0.0076**, **WER 0.0261**, **BLEU 0.9396**, **ANLS 0.9920**, and **METEOR 0.9723**. On **Tamil**, its performance is much weaker: **CER 0.1392**, **WER 0.6500**, **BLEU 0.1487**, **ANLS 0.8672**, and **METEOR 0.3359**. The comparison is asymmetric. Surya strongly outperforms **Cloud Vision API**, **Document AI**, **Tesseract 5.5.0**, and **Subasa OCR** on Sinhala, but is far behind **Document AI** and **Cloud Vision API** on Tamil, and is also worse than **EasyOCR** on Tamil in the core OCR metrics. The paper interprets this as evidence that multilingual zero-shot transfer for low-resource scripts remains highly language-dependent even when scripts are typologically close [2507.18264].

A second evaluation, **KITAB-Bench**, places Surya in a broader Arabic OCR and document-understanding benchmark. There Surya is tested on **line detection**, **line recognition**, **layout detection**, **table extraction**, **OCR/image-to-text**, and **end-to-end PDF-to-Markdown**. The benchmark contains **8,809 samples** across **9 major domains** and **36 sub-domains**. Surya’s profile is mixed. On the full **3,760-image OCR set**, it performs very poorly for Arabic recognition, with **CER 4.95** and **WER 5.61**. On line detection, it is substantially stronger, with **mAP@50 = 79.67** and **mAP@0.5:0.95 = 27.40**; on line recognition the reported values are **WER = 1.01** and **CER = 0.87**. In layout detection it is not the top detector by mAP, but it has relatively strong precision: **0.751** on BCE and **0.782** on DocLayNet. In table extraction it obtains **TEDS 50.15** and **Jaccard 70.42**; for end-to-end PDF-to-Markdown it obtains **chrF 58.38**, **TEDS 44.29**, and **MARS 51.34** [2502.14949].

The Arabic benchmark’s interpretation is not that Surya is uniformly weak, but that it is more competitive in **structural tasks** than in **Arabic text recognition itself**. The benchmark attributes Arabic difficulty to **cursive script**, **right-to-left flow**, **complex typography/calligraphy**, **font diversity**, **diacritical marks**, **numeral recognition**, **word elongation**, and structural issues such as **merged cells**, **tables**, and **layout complexity**. A plausible implication is that Surya’s architecture transfers more effectively to geometric and structural subtasks than to fine-grained script-specific transcription under Arabic document conditions [2502.14949].

## 3. Surya as a heliophysics foundation model

In “Surya: Foundation Model for Heliophysics,” Surya denotes a **foundation model for heliophysics** rather than an OCR system. It is introduced as a **366M parameter** model trained on multi-instrument **Solar Dynamics Observatory (SDO)** observations, specifically **eight Atmospheric Imaging Assembly (AIA) channels** and **five Helioseismic and Magnetic Imager (HMI) products**, for a total of **13 channels** per observation. Each observation has the form
$$
X \in \mathbb{R}^{13 \times 4096 \times 4096},
$$
and the model uses **two timestamps** as input, giving
$$
13 \times 2 \times 4096 \times 4096.
$$
The processed dataset spans **May 13, 2010 to December 31, 2024**, with about **257 TB** of processed data and a standardized cadence of **12 minutes** [2508.14112].

Architecturally, Surya is a **2-D spatiotemporal transformer** with **spectral gating blocks**, **long--short range attention blocks**, and a decoder that maps token features back to image space. Tokenization uses patch size \(16 \times 16\), hidden dimension \(D = 1280\), and **65,536** total tokens, together with **Fourier position embeddings**. The spectral gating component applies a real FFT, learnable complex-valued frequency weights, and inverse FFT:
$$
\widetilde{\mathbf{X}} = \mathcal{F}(\mathbf{X}_s), \qquad
\widetilde{\mathbf{X}}' = \widetilde{\mathbf{X}} \odot W_c, \qquad
\mathbf{X}_s' = \mathcal{F}^{-1}(\widetilde{\mathbf{X}}'),
$$
followed by residual refinement,
$$
\mathbf{X}_{\ell+1} = \mathbf{X}_\ell + \mathrm{FFN}\!\left(\mathrm{LN}(\mathbf{X}')\right).
$$
Its long--short design combines local windowed attention and efficient global aggregation, reflecting the paper’s claim that solar forecasting requires both fine active-region detail and global coronal context [2508.14112].

Pretraining is organized around **time advancement / forecasting** rather than masked reconstruction. The model is given **two timestamps 60 minutes apart** and trained to predict the SDO state **60 minutes in the future** with
$$
\mathcal{L}_{\text{MSE}} = \frac{1}{N}\left\lVert \mathbf{X}_{t+1} - f_\theta(\mathbf{X}_t,\mathbf{X}_{t-1}) \right\rVert^2.
$$
A second stage performs **autoregressive rollout tuning**, including
$$
\mathcal{L}_{\text{rollout}} = \frac{1}{2N}\Big(
\left\lVert \mathbf{X}_{t+1} - f_\theta(\mathbf{X}_t,\mathbf{X}_{t-1}) \right\rVert^2 +
\left\lVert \mathbf{X}_{t+2} - f_\theta(f_\theta(\mathbf{X}_t,\mathbf{X}_{t-1}), \mathbf{X}_t) \right\rVert^2
\Big).
$$
The rollout schedule is **160,000 steps** for one-step training, then **20,000 steps** for 2-step ahead, and **4,000 steps each** for **3-, 4-, 5-step ahead** tuning. The implementation relies on **FSDP**, **mixed precision**, and **gradient checkpointing** because of the memory demands of full-disk \(4096 \times 4096\) inputs [2508.14112].

Zero-shot and downstream results position Surya as a general-purpose solar representation learner. The paper reports that rollout tuning improves **12-hour-ahead** performance by up to **17.8%**, and that the model can be tuned to **24-hour lead times** on an **80 GB A100**. For downstream tasks, adaptation uses **LoRA** with
$$
W = W_0 + \frac{\alpha}{r}BA.
$$
Reported results include **IoU 0.768** and **Dice 0.853** for active-region segmentation with only **4.1M trainable parameters**; **TSS 0.436**, **HSS 0.522**, and **F1 0.561** for 24-hour **M- or X-class flare** forecasting; **MSE 0.000126**, **MAE 0.00451**, and **MAPE 1.48%** for **1343-band solar EUV irradiance** prediction; and **RMSE 75.92 km/s** with **MAE 58.06 km/s** for **4 days ahead** solar wind forecasting. The paper also notes a limitation: because the model is deterministic and trained with MSE, it tends to **blur sharp features**, even though it may denoise low-SNR channels such as **94 Å** and **335 Å** [2508.14112].

## 4. Surya as author surname in theoretical physics and discrete spacetime

In several papers, **Surya** is part of an author’s name rather than a model designation. The clearest example is “The Emergence of Spectral Universality in Deep Networks,” where the paper’s content concerns the singular value spectrum of a deep network’s input-output Jacobian at initialization, and “Surya” refers only to **Surya Ganguli**. The paper studies fully connected networks
$$
\mathbf{x}^l = \phi(\mathbf{h}^l), \qquad \mathbf{h}^l = \mathbf{W}^l \mathbf{x}^{l-1} + \mathbf{b}^l,
$$
with Jacobian
$$
\mathbf{J} = \frac{\partial \mathbf{x}^L}{\partial \mathbf{x}^0} = \prod_{l=1}^L \mathbf{D}^l \mathbf{W}^l,
$$
and analyzes \(\mathbf{J}\mathbf{J}^T\) using **free probability theory** and the **S-transform**. A central scalar is
$$
\chi = \sigma_w^2 \int \mathcal{D}h\;\big[\phi'(\sqrt{q^*}h)\big]^2,
$$
which locates the critical line \(\chi=1\), but the paper emphasizes that criticality alone is not sufficient because the full spectrum may still broaden with depth. Its practical conclusion is that **orthogonal weights** are necessary for stable limiting distributions, and that **ReLU** does not preserve the same favorable spectral behavior at large depth as some alternatives such as **Hard Tanh**, **Erf**, and **SiLU** [1802.09979].

In mathematical relativity, Surya appears through the **Page–Surya–Woolgar positive mass theorem**, which motivates the rigidity problem studied in “Conformal boundary rigidity from null geodesic travel times.” That paper asks whether an asymptotically anti-de Sitter spacetime is conformally AdS if every null geodesic from a boundary point returns to the **spacetime antipode** at conformal infinity. The paper proves affirmative rigidity results in three regimes: under the **null energy condition**, in the **globally static** case, and in the **globally stationary** case. The stationary proof uses the quotient optical manifold, where null geodesics project to **magnetic geodesics** with Mañé action
$$
A(\gamma)=L_{\tilde h}(\gamma)-\int_\gamma \omega.
$$
Its main stationary theorem states that if every such null geodesic passes through the antipode, then the spacetime is **conformally isometric to anti-de Sitter spacetime** [2606.28689].

In causal set theory, the paper “On the Axioms of Causal Set Theory” treats **Sumati Surya** primarily as a major review source for causal set phenomenology and as part of the literature on maximal antichains. It cites Surya’s review *Directions in Causal Set Quantum Gravity* as a “fine overview” and an “excellent starting point” for topological issues. The paper then argues for revisions to standard causal set axioms, including replacing **transitivity** by a more primitive **causal preorder**, replacing **interval finiteness** by **local finiteness**, and moving from element-space antichains to **relation space**, where maximal antichains are claimed to be **impermeable**. The critique is explicitly framed against the background of ordinary causal set theory and the literature involving **Major, Rideout, and Surya** on the permeability of maximal antichains [1311.2148].

## 5. Surya in stochastic processes, combinatorics, and mathematical finance

In probability and stochastic-process theory, Surya is associated with a sequence of papers on **mixtures of finite-state Markov jump processes**. The 2018 and 2018 follow-up papers summarized in the data treat **Surya (2018)** as the precursor that extended Frydman-type mixtures to absorbing continuous-time Markov jump processes. Subsequent work generalizes this framework to **conditional joint distributions of first exit times**, **overlapping absorbing sets**, explicit **Bayesian updates** for latent regime probabilities and state probabilities, and reductions to classical **phase-type** laws when all regimes move at the same speed. These papers emphasize that the observed mixture is non-Markovian, while still admitting explicit formulas for conditional transition matrices, survival functions, densities, Laplace transforms, and moments in terms of intensity matrices and posterior regime weights [1809.06713] [1807.04591].

The later paper “A new class of conditional Markov jump processes with regime switching and path dependence” is explicitly positioned as an extension of **Budhi Surya’s** earlier work with Frydman. Its main novelty is to reinterpret a finite mixture of Markov jump processes as a **distributionally equivalent conditional Markov jump process** that **reselects the latent speed regime at each jump epoch** according to the current state, current time, and path history. This produces explicit formulas for conditional holding times and transition laws, together with a full inference pipeline based on closed-form **MLEs**, an **EM algorithm**, **Akaike information criterion**, an explicit observed Fisher information matrix in a simplified **Louis (1982)** form, and asymptotic results including consistency, asymptotic normality, and efficiency properties relative to the **Cramér–Rao** bound [2107.07026].

In combinatorics, Surya appears in the phrase **Spiro and Surya**, whose earlier univariate result on balanced complete \(r\)-partite graphs is extended in “Deranged Perfect Matchings on complete graph and balanced complete r-partite graph.” The earlier result proved that when \(R\) is a uniformly random perfect matching of \(K_{r \times 2n/r}\), the number of edges of \(R\) lying in a fixed balanced perfect matching \(M\) converges to a Poisson law with mean \(r/(2r-2)\). The 2025 paper generalizes this to the multivariate setting, obtaining asymptotic independent Poisson limits for vectors of edge-intersection counts in sparse subgraphs [2505.13799].

In mathematical finance, Surya appears through **Kyprianou and Surya (2007)** in endogenous-bankruptcy models with **spectrally negative Lévy** asset dynamics. “Optimal Capital Structure with Scale Effects under Spectrally Negative Levy Models” treats that earlier work as its immediate predecessor and generalizes it by allowing **state-dependent bankruptcy costs** and **state-dependent tax benefits**. The resulting model preserves threshold-type bankruptcy and uses **scale functions** \(W^{(q)}\), \(Z^{(q)}\), and \(\Theta^{(q)}\) to derive a candidate optimal threshold and a sufficient condition for optimality. The paper’s economic sufficient condition is that \(\eta\) is increasing, \(\overline\eta\) is decreasing, \(f\) is increasing, and \(0 \le \overline\eta(\cdot) \le 1\), which encodes the stated **scale effects** [1109.0897].

## 6. Surya as the referent of Surya Namaskar in pose-recognition work

In “Surya Namaskar: real-time advanced yoga pose recognition and correction for smart healthcare,” the term appears in its religious-cultural sense. **Suryanamaskar**, also described as **salute to the sun**, is a yoga practice devoted to the Hindu Sun God, **Surya**. The paper proposes a real-time recognition-and-correction system in which a user performs Surya Namaskar in front of a camera, **MediaPipe Holistic** extracts body landmarks, the landmarks are converted into **NumPy** arrays, and a deep model built with **TensorFlow/Keras** classifies the observed sequence into one of **8 Surya Namaskar pose classes**: **Pranamasana**, **Hasta Uttanasana**, **Hasta Padasana**, **Ashwa-Sanchalanasana**, **Ashtanga Namaskara**, **Dandasana**, **Bhujangasana**, and **Svanasana** [2209.02492].

The dataset is **manually created by the authors** from videos recorded **inside a room**, with a camera **several meters away**, at **60 fps**, and from **front, back, and side** views. Although the abstract mentions **CNN**, the architecture table is a **stacked LSTM network** with output shapes `(None, 10, 64)`, `(None, 10, 128)`, `(None, 64)`, followed by dense layers of **64**, **32**, and **8** units, for a total of **596,840 trainable parameters**. The paper reports two numerical results: **model accuracy of 98.68 percent** and `accuracy_score(ytrue, yhat) = 0.75`. It does not reconcile these values fully, and the data explicitly notes that the reporting is somewhat inconsistent. The intended functionality is a **real-time pose recognition and correction assistant**, though the paper does not provide a sophisticated rule-based correction engine with joint-angle thresholds [2209.02492].

Across these literatures, the principal encyclopedic point is therefore disambiguation. **Surya OCR** and **Surya for heliophysics** are model names with concrete architectures, datasets, training objectives, and benchmark suites; **Surya Ganguli**, **Sumati Surya**, and **Budhi Surya** are authors whose names appear in technically unrelated lines of work; and **Surya Namaskar** preserves the older religious referent of the term. Any technical reading of “Surya” in arXiv materials is incomplete unless the relevant paper is identified first.

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