Beamer Frankfurt

\documentclass[10pt]{beamer}
\usetheme{Frankfurt}
\usecolortheme{whale}
\usepackage[utf8]{inputenc}
\usepackage[T1]{fontenc}
\usepackage{amsmath,amssymb}
\usepackage{graphicx}
\usepackage{booktabs}
\usepackage{hyperref}

\title[Graph Learning]{Graph Neural Networks for Scientific Discovery}
\subtitle{Methods and Open Problems}
\author{First Last}
\institute[Example Labs]{Example Research Laboratory}
\date{\today}

\begin{document}

\frame{\titlepage}
\begin{frame}{Outline}\tableofcontents\end{frame}

\section{Introduction}
\begin{frame}{Why Graphs?}
\begin{itemize}
  \item Many scientific domains are naturally graph-structured.
  \item Molecules, proteins, materials, citation networks.
  \item Standard neural networks cannot exploit this structure.
\end{itemize}
\end{frame}

\begin{frame}{Motivating Example}
Predicting molecular properties: atoms are nodes, bonds are edges.
A good model must be permutation-invariant over atom labelings.
\end{frame}

\section{Prior Work}
\begin{frame}{Classical Approaches}
\begin{itemize}
  \item Hand-crafted molecular fingerprints.
  \item Kernel methods on graphs.
  \item Random-walk embeddings.
\end{itemize}
\end{frame}

\begin{frame}{Recent Progress}
Message-passing neural networks, graph convolutions, graph transformers have
pushed performance on many benchmarks in the past five years.
\end{frame}

\section{Our Approach}
\begin{frame}{Equivariant Message Passing}
\begin{itemize}
  \item Nodes update by aggregating neighbor features.
  \item Aggregation is permutation-invariant.
  \item We add an equivariant readout for 3D geometry.
\end{itemize}
\end{frame}

\begin{frame}{Training Objective}
\[ \mathcal{L} = \frac{1}{N}\sum_i \|\hat y_i - y_i\|^2 + \lambda\|W\|_F^2 \]
We use a standard Adam optimizer with cosine warm-up.
\end{frame}

\section{Experiments}
\begin{frame}{Benchmarks}
\begin{table}
\centering
\begin{tabular}{lcc}
\toprule
Task & Prior SoTA & Ours \\
\midrule
QM9 ($U_0$, meV)         & 11.0 & \textbf{8.3} \\
OC20 (eV)                & 0.28 & \textbf{0.21} \\
MatBench (ROC-AUC)       & 0.86 & \textbf{0.90} \\
\bottomrule
\end{tabular}
\end{table}
\end{frame}

\begin{frame}{Ablations}
Removing equivariance hurts accuracy by 18\%; removing the skip connections
hurts it by another 7\%. Both matter; equivariance more.
\end{frame}

\section{Closing}
\begin{frame}{Open Problems}
\begin{itemize}
  \item Scaling beyond 100k atoms remains memory-bound.
  \item Interpretability tools are immature.
  \item Transfer across chemistries is still an open question.
\end{itemize}
\end{frame}

\begin{frame}{Thank You}
\centering
\Huge Questions?\\[0.8em]
\normalsize [email protected]
\end{frame}

\end{document}