% PERELMAN — AI4X-AC 2026 Poster
% Conference: AI for Accelerated Science (AI4X-AC 2026), Singapore, June 15–19
% Submission #316 | Accept (Poster)
% Format: A0 Portrait (841×1189 mm) — standard academic poster
%
% Compile: pdflatex perelman_ai4xac2026_poster.tex
% Requirements: beamerposter, tikz, booktabs, graphicx, amsmath, fontenc, inputenc, xcolor

\documentclass[final]{beamer}

\usepackage[T1]{fontenc}
\usepackage[utf8]{inputenc}
\usepackage{lmodern}
\usepackage[stretch=10,shrink=10]{microtype}
\usepackage{graphicx}
\usepackage{booktabs}
\usepackage{amsmath,amssymb}
\usepackage{xcolor}
\usepackage{tikz}
\usepackage{multicol}
\usepackage{array}
\usepackage{tabularx}
\usepackage{hyperref}
\usepackage{fontawesome5}

%% --- Color scheme ---
\definecolor{ai4xblue}{RGB}{20, 74, 153}     % Deep blue (AI4X-AC brand)
\definecolor{ai4xlightblue}{RGB}{230, 238, 255}
\definecolor{ai4xaccent}{RGB}{0, 153, 102}   % Green accent
\definecolor{ai4xwarn}{RGB}{200, 80, 0}      % Orange for highlights
\definecolor{sectionbg}{RGB}{240, 245, 255}
\definecolor{lightgray}{RGB}{248, 248, 248}

%% --- Beamer theme ---
\usetheme{default}
\setbeamercolor{headline}{fg=white, bg=ai4xblue}
\setbeamercolor{footline}{fg=white, bg=ai4xblue}
\setbeamercolor{block title}{fg=white, bg=ai4xblue}
\setbeamercolor{block body}{fg=black, bg=sectionbg}
\setbeamercolor{block title alerted}{fg=white, bg=ai4xwarn}
\setbeamercolor{block body alerted}{fg=black, bg=white}
\setbeamerfont{block title}{size=\large, series=\bfseries}
\setbeamerfont{headline}{size=\Large}

%% --- Poster size: A0 portrait ---
\usepackage[orientation=portrait, size=a0, scale=1.05]{beamerposter}

%% === Helper commands ===
\newcommand{\boldblue}[1]{\textcolor{ai4xblue}{\textbf{#1}}}
\newcommand{\boldgreen}[1]{\textcolor{ai4xaccent}{\textbf{#1}}}
\newcommand{\boldred}[1]{\textcolor{ai4xwarn}{\textbf{#1}}}
\newcommand{\metric}[2]{%
  \begin{center}
    {\fontsize{36}{40}\selectfont\textbf{\textcolor{ai4xblue}{#1}}}\\[0.3em]
    {\large\textcolor{black}{#2}}
  \end{center}
}

%% === Document ===
\begin{document}

%% ============================================================
%% HEADLINE
%% ============================================================
\begin{frame}[t]{}
\vspace{-2cm}

\begin{beamercolorbox}[wd=\textwidth, ht=10cm, center, dp=1cm]{headline}
  \vspace{1cm}
  {\fontsize{52}{60}\selectfont\textbf{PERELMAN}} \\[0.5em]
  {\fontsize{28}{34}\selectfont Automated Quantitative Meta-Analysis of Scientific Literature}\\
  {\fontsize{28}{34}\selectfont via Vision-Language Models}\\[1em]
  {\fontsize{20}{26}\selectfont
    Daniil Merkulov\textsuperscript{1,2} \quad Danil Sherki\textsuperscript{2} \quad [co-authors TBD]
  }\\[0.5em]
  {\fontsize{18}{24}\selectfont
    \textsuperscript{1}Skoltech \quad \textsuperscript{2}AI for Science Lab, Sber \quad
    \texttt{d.merkulov@skoltech.ru}
  }
\end{beamercolorbox}

\vspace{0.5cm}

%% ============================================================
%% MAIN BODY — 3-column layout
%% ============================================================
\begin{columns}[T]

%% ====== LEFT COLUMN ======
\begin{column}{0.31\textwidth}

  % ----- PROBLEM -----
  \begin{block}{\faSearch\quad The Data Extraction Bottleneck}
    \vspace{0.3em}
    Scientific literature grows at $>$2\,M papers/year. For materials
    science, drug discovery, and physics, critical quantitative results
    are \textbf{not in tables or abstracts — they live in figures}.

    \vspace{0.8em}
    \begin{alertblock}{Motivating Example}
      A systematic review of \textbf{NMC811 cathode materials}
      (Savina \& Abakumov) required processing \boldblue{548 papers}
      with \boldblue{950+ experiments} — months of manual work for a PhD student.
    \end{alertblock}

    \vspace{0.8em}
    \textbf{Key insight:} $\sim$40--60\% of quantitative results
    appear \textit{only in figures}, not in text or tables.
    Text-based retrieval methods (Elicit, SciSpace) miss this data entirely.

    \vspace{0.8em}
    \includegraphics[width=0.95\linewidth]{poster_figures/fig4_pipeline_diagram.png}
    \begin{center}
    {\small \textit{Fig.~1: PERELMAN pipeline — from PDF corpus to structured database}}
    \end{center}
  \end{block}

  \vspace{0.5em}

  % ----- COMPETITIVE LANDSCAPE -----
  \begin{block}{\faChartBar\quad Competitive Landscape}
    \vspace{0.3em}
    {\footnotesize
    \begin{tabularx}{0.98\linewidth}{@{}lXll@{}}
      \toprule
      \textbf{System} & \textbf{Task} & \textbf{Expert?} & \textbf{Output} \\
      \midrule
      \boldblue{PERELMAN} & \boldblue{Meta-analysis corpus} & \boldblue{No} & \boldblue{Struct.\ DB} \\
      AutoResearchClaw & Write a paper & Topic only & Draft paper \\
      Elicit/SciSpace & Q\&A over abstracts & No & Text answers \\
      Manual review & Meta-analysis & Full process & Tables \\
      \bottomrule
    \end{tabularx}
    }

    \vspace{0.8em}
    \textbf{Unique niche:} PERELMAN extracts \textit{numbers from graphs}
    — a capability that no text-based RAG can replicate.
  \end{block}

  \vspace{0.5em}

  % ----- APPLICATIONS -----
  \begin{block}{\faFlask\quad Applications}
    \vspace{0.3em}
    \begin{itemize}
      \item \textbf{Materials Science:} capacity/voltage DB from battery papers
        $\to$ regression models linking synthesis params to performance
      \item \textbf{Drug Discovery:} IC50/EC50 from pharmacology graphs
      \item \textbf{Physics:} critical temperatures, band gaps, phase diagrams
      \item \textbf{General pattern:} any domain with characteristic graph types
        $\to$ one domain-specific prompt $\to$ process entire corpus
    \end{itemize}

    \vspace{0.8em}
    \begin{alertblock}{Cost estimate}
      \$0.01--0.05 per paper (VLM calls via OpenRouter). \\
      548 papers $\approx$ \$5--25 total — negligible vs.\ months of manual annotation.
    \end{alertblock}
  \end{block}

\end{column}

%% ====== MIDDLE COLUMN ======
\begin{column}{0.36\textwidth}

  % ----- METHOD -----
  \begin{block}{\faCogs\quad Pipeline Overview}
    \vspace{0.3em}
    PERELMAN is a multi-stage VLM pipeline for automated figure-based data extraction:

    \vspace{0.5em}
    \begin{center}
    \begin{tikzpicture}[
        box/.style={draw=ai4xblue, fill=ai4xlightblue, rounded corners=6pt,
                    minimum width=10cm, minimum height=1.2cm,
                    text width=9.5cm, align=center, font=\normalsize},
        arrow/.style={->, thick, color=ai4xblue},
        annot/.style={text width=8cm, align=left, font=\small\itshape, text=gray}
      ]
      \node[box] (pdf)   at (0,0)     {\textbf{PDF Corpus} (548 papers)};
      \node[box] (doc)   at (0,-2.0)  {\textbf{docling} — PDF $\to$ Markdown + figures};
      \node[box] (mos)   at (0,-4.0)  {\textbf{Mosaic Detector} (VLM)\\
                                       {\small rows$\times$cols, panel labels, target panel}};
      \node[box] (spl)   at (0,-6.0)  {\textbf{Panel Splitter + Score Ranker}\\
                                       {\small caption context $\pm$1000 chars}};
      \node[box] (ext)   at (0,-8.0)  {\textbf{VLM Extractor} (target panel only)\\
                                       {\small voltage\_range, C\_dis, C\_ch, CE\%}};
      \node[box] (txt)   at (0,-10.0) {\textbf{Text Fallback} (LLM on paper text)};
      \node[box] (json)  at (0,-12.0) {\textbf{Sidecar JSON} — Structured Database};

      \draw[arrow] (pdf)  -- (doc);
      \draw[arrow] (doc)  -- (mos);
      \draw[arrow] (mos)  -- (spl);
      \draw[arrow] (spl)  -- (ext);
      \draw[arrow] (ext)  -- (txt);
      \draw[arrow] (txt)  -- (json);
    \end{tikzpicture}
    \end{center}

    \vspace{0.5em}
    \textbf{Key innovations:}
    \begin{itemize}
      \item \textbf{docling:} OCR + structure preservation (better than PyMuPDF)
      \item \textbf{Mosaic Detector:} custom VLM prompt for multi-panel figures (a/b/c/d)
      \item \textbf{Context-aware prompting:} $\pm$1000\,char window around figure references
        (caption + in-text mentions) — improves VLM extraction precision
      \item \textbf{Zero-shot domain adaptation:} NMC811 synonyms + baseline-series logic
    \end{itemize}
  \end{block}

  \vspace{0.5em}

  % ----- ACCURACY PROGRESS -----
  \begin{block}{\faChartLine\quad Accuracy Progression}
    \vspace{0.3em}
    \begin{center}
    \begin{tikzpicture}
      \draw[->, thick] (0,0) -- (9,0) node[right] {\small Version};
      \draw[->, thick] (0,0) -- (0,4.5) node[above] {\small Accuracy (\%)};
      % v0
      \filldraw[ai4xwarn] (1,2.0) circle(8pt) node[above=5pt]{\small v0: 50\%};
      % v1
      \filldraw[ai4xblue!60] (4,3.2) circle(8pt) node[above=5pt]{\small v1: 71\%};
      % v2
      \filldraw[ai4xaccent] (7,3.7) circle(8pt) node[above=5pt]{\small v2: 83\%};
      \draw[thick, dashed, gray] (1,2.0) -- (4,3.2) -- (7,3.7);
      % labels
      \node[font=\footnotesize, gray] at (1,-0.5) {baseline};
      \node[font=\footnotesize, gray] at (4,-0.5) {+context};
      \node[font=\footnotesize, gray] at (7,-0.5) {+mosaic};
      % grid
      \foreach \y/\lab in {1/25, 2/50, 3/75, 4/100}{
        \draw[gray!30] (0,\y) -- (8.5,\y);
        \node[left, font=\footnotesize, gray] at (0,\y) {\lab};
      }
    \end{tikzpicture}
    \end{center}

    \vspace{0.3em}
    Each ablation stage adds a key module:
    v0 = direct VLM on figure; v1 = + context-aware prompting (+21pp);
    v2 = + mosaic detection and baseline-series logic (+12pp).
  \end{block}

\end{column}

%% ====== RIGHT COLUMN ======
\begin{column}{0.31\textwidth}

  % ----- RESULTS -----
  \begin{block}{\faTrophy\quad Results on NMC811 Corpus}
    \vspace{0.3em}

    \textbf{Evaluation:} 24 papers, ground truth from Savina \& Abakumov.

    \vspace{0.8em}
    % Big metrics
    \begin{columns}[T]
      \begin{column}{0.48\linewidth}
        \metric{4.6\%}{Capacity MAPE\\(n=20 matched pairs)}
      \end{column}
      \begin{column}{0.48\linewidth}
        \metric{83\%}{Coverage\\(20/24 papers)}
      \end{column}
    \end{columns}

    \vspace{0.5em}
    \begin{columns}[T]
      \begin{column}{0.48\linewidth}
        \metric{95\%}{Voltage max\\accuracy ($\pm$0.1V, n=21)}
      \end{column}
      \begin{column}{0.48\linewidth}
        \metric{86\%}{Voltage min\\accuracy ($\pm$0.1V, n=21)}
      \end{column}
    \end{columns}

    \vspace{0.8em}
    \includegraphics[width=0.95\linewidth]{poster_figures/fig1_capacity_scatter.png}
    \begin{center}
    {\small \textit{Fig.~2: Extracted vs.\ ground-truth capacity (n=20, MAPE=4.6\%)}}
    \end{center}

    \vspace{0.5em}
    \includegraphics[width=0.95\linewidth]{poster_figures/fig3_accuracy_bars.png}
    \begin{center}
    {\small \textit{Fig.~3: Accuracy by metric — voltage extraction is highly precise}}
    \end{center}
  \end{block}

  \vspace{0.5em}

  % ----- ERROR ANALYSIS -----
  \begin{block}{\faSearchMinus\quad Error Analysis}
    \vspace{0.3em}
    \includegraphics[width=0.95\linewidth]{poster_figures/fig2_error_distribution.png}
    \begin{center}
    {\small \textit{Fig.~4: Error distribution — median 5.8\,mAh/g, 9/20 papers $<$2\,mAh/g error}}
    \end{center}

    \vspace{0.5em}
    \textbf{Failure modes (17\% miss):}
    \begin{itemize}
      \item Non-standard panel layouts (3-column, overlapping)
      \item Extremely low-resolution figures
      \item Papers reporting only normalized capacity (C/C$_0$)
    \end{itemize}
  \end{block}

  \vspace{0.5em}

  % ----- CONCLUSION -----
  \begin{block}{\faCheckCircle\quad Conclusion}
    \vspace{0.3em}
    PERELMAN reduces quantitative literature curation from
    \textbf{months to hours}, enabling data-driven modeling at the scale
    of entire research domains.

    \vspace{0.5em}
    \begin{itemize}
      \item \boldgreen{4.6\% MAPE} — publication-quality precision
      \item \boldgreen{83\% coverage} — high throughput on real corpus
      \item \boldgreen{Domain-agnostic architecture} — adapt with one prompt
      \item \boldgreen{Open pipeline} — docling + OpenRouter VLMs
    \end{itemize}

    \vspace{0.5em}
    \textbf{Future work:} generalization to drug discovery / physics corpora;
    active learning for prompt refinement; integration with
    downstream regression models.

    \vspace{0.3em}
    \textbf{Demo dataset:} NMC811 database (548 papers, 950+ experiments)
    available on request.
  \end{block}

  \vspace{0.5em}

  % ----- REFERENCES (condensed) -----
  \begin{block}{\faBookOpen\quad Key References}
    \vspace{0.3em}
    {\footnotesize
    \begin{itemize}
      \item Savina \& Abakumov (2024). NMC811 systematic review. \textit{[corpus source]}
      \item Auer et al. (2024). docling — Document Understanding \& Conversion. arXiv:2408.09869
      \item Submission \#316, AI4X-AC 2026, Singapore
    \end{itemize}
    }
  \end{block}

\end{column}

\end{columns}

\vspace{0.2cm}

%% ============================================================
%% FOOTLINE
%% ============================================================
\begin{beamercolorbox}[wd=\textwidth, ht=2.0cm, center, dp=0.3cm]{footline}
  \vspace{0.2cm}
  \begin{minipage}[c]{0.19\linewidth}\centering
    {\small Skoltech / AI4Science, Sber}
  \end{minipage}\hfill
  \begin{minipage}[c]{0.58\linewidth}\centering
    {\small\textbf{AI for Accelerated Science 2026 (AI4X-AC)}}\\
    {\small Singapore · June 15--19 \quad · \quad Poster Session · Submission \#316}
  \end{minipage}\hfill
  \begin{minipage}[c]{0.19\linewidth}\centering
    {\small \texttt{d.merkulov@skoltech.ru}}
  \end{minipage}
\end{beamercolorbox}

\end{frame}
\end{document}